Automatic measurement of lesions on medical images

Abstract

The present invention relates to a method for assessing the presence and/or the severity of a lesion in an organ or tissue of a subject through automated analysis of at least one image of said organ or tissue, wherein said organ or tissue is preferably a liver organ or liver tissue, comprising the calculation of a score combining descriptors of said image, wherein said method comprises the steps of: a. measuring on said at least one image at least two descriptors of said at least one image; b. mathematically combining said at least two descriptors in a score; and c. assessing the presence and/or the severity of a lesion in the organ or tissue based on the value of the score calculated at step (b).

Claims

1. A non-invasive method for assessing the presence of a clinically significant fibrosis in a liver organ or liver tissue of a subject through automated analysis of at least one image of said liver organ or liver tissue, the method utilizing a microprocessor comprising a computer algorithm configured to calculate a score combining descriptors of said image, said method comprising the steps of: a. measuring on said at least one image at least two descriptors of said at least one image, the descriptors selected from the group consisting of: linearity percentage of the edges, mean of percentage of fibrosis around areas, area of stellar fibrosis among the total surface of the liver biopsy specimen, number of bridges, bridges thickness, mean area of porto-septal regions, bridges perimeter, ratio of bridges among the porto-septal areas, area of fibrosis in the bridges, fractal dimension of peri-sinusoidal fibrosis, perimeter of the liver organ, liver tissue or fragment thereof, fractal dimension of porto-septal fibrosis, ratio of peri-sinusoidal fibrosis among the whole fibrosis, length of the liver organ, liver tissue or fragment thereof, anfractuosity descriptors selected from native perimeter, smoothed perimeter and ratio between both perimeters, fractal dimension of fibrosis, interquartile range of total density, Arantius furrow thickness, mean native liver perimeter, mean total spleen perimeter, ratio spleen surface to liver surface, and mathematic combination thereof; b. mathematically combining said at least two measured descriptors from step (a) in to a score, wherein said mathematical combination is a binary logistic regression; and c. assessing the presence of a clinically significant fibrosis, based on the value of the score calculated at step (b).

2. The method according to claim 1, wherein the at least one image is a histological section image, or an image of the entire liver organ or liver tissue.

3. The method according to claim 1, wherein the at least one image is obtained by an optical technique selected from the group consisting of: microscopic physical imaging, second harmonic generation (SHG), multiphoton imaging, coherent anti-Stokes Raman scattering (CARS), two-photon excitation fluorescence (TPEF), diffuse optical imaging and event-related optical signal; or by a non-optical technique selected from the group consisting of: radiography, nuclear medicine, photoacoustic methods, and thermal methods.

4. The method according to claim 1, further comprising measuring at least one non-invasive test marker, measuring liver stiffness, or measuring both, in the subject and combining the resulting measure(s) with the measures of the two descriptors obtained at step (a).

5. The method according to claim 4, wherein the at least one non-invasive test marker is selected from the group consisting of a biomarker, a clinical data, a physical data, and a score selected from Enhanced Liver Fibrosis test (ELF™) FibroSpect™, AST to Platelet Ratio Index (APRI), Fibrosis-4 (FIB-4), Hepascore, Fibrotest™, FibroMeter™, CirrhoMeter™, CombiMeter™, and InflaMeter™.

6. The method according to claim 1, wherein the clinically significant fibrosis is proto-septal fibrosis.

7. A non-invasive method for assessing the presence of cirrhosis in a subject through automated analysis of at least one image of liver organ or liver tissue from the subject, the method utilizing a microprocessor comprising a computer algorithm configured to calculate a score combining descriptors of said image, said method comprising the steps of: a. measuring on said at least one image at least two descriptors of said at least one image, the descriptors selected from the group consisting of: linearity percentage of the edges, granularity percentage, fragmentation, mean of percentage of fibrosis around areas, area of stellar fibrosis among the surface of porto-septal regions, portal distance, length of the liver biopsy, fractal dimension of porto-septal fibrosis, number of bridges, fractal dimension of peri-sinusoidal fibrosis, perimeter of the liver organ, liver tissue or fragment thereof, standard deviation of total density, second antero-posterior segment I length (D3), Arantius furrow thickness, ratio between spleen surface and liver surface, total spleen perimeter, portal furrow thickness and mathematic combination thereof; b. mathematically combining said at least two measured descriptors from step (a) in to a score, wherein said mathematical combination is a binary logistic regression; and c. assessing the presence of a cirrhosis, based on the value of the score calculated at step (b).

8. The method according to claim 7, wherein the at least one image is a histological section image, or an image of the entire liver organ or liver tissue.

9. The method according to claim 7, wherein the at least one image is obtained by an optical technique selected from the group consisting of: microscopic physical imaging, second harmonic generation (SHG), multiphoton imaging, coherent anti-Stokes Raman scattering (CARS), two-photon excitation fluorescence (TPEF), diffuse optical imaging and event-related optical signal; or by a non-optical technique selected from the group consisting of: radiography, nuclear medicine, photoacoustic methods, and thermal methods.

10. The method according to claim 7, further comprising measuring at least one non-invasive test marker selected from the group consisting of a biomarker, a clinical data, a physical data and a score selected from Enhanced Liver Fibrosis test (ELF™), FibroSpect™, AST to Platelet Ratio Index (APRI), Fibrosis-4 (FIB-4), Hepascore, Fibrotest™, FibroMeter™, CirrhoMeter™, CombiMeter™, InflaMeter™, and liver stiffness, and combining the resulting measure(s) with the measures of the two descriptors obtained at step (a).

11. A non-invasive method for determining an increased risk of liver complications, an increased risk of mortality, or both, through automated analysis of at least one image of liver organ or liver tissue, the method utilizing a microprocessor comprising a computer algorithm configured to calculate a score combining descriptors of said image, wherein said method comprises the steps of: a. measuring on said at least one image at least two descriptors of said at least one image, the descriptors selected from the group consisting of: linearity percentage of the edges, granularity percentage, fragmentation, mean of percentage of fibrosis around areas, area of stellar fibrosis among the surface of porto-septal regions, area of stellar fibrosis among the surface of lobular regions, area of porto-septal fibrosis, portal distance, fractal dimension of the edges of the organ, tissue or fragment thereof, area of steatosis, mean intensity of the image on the blue component, number of nodules, number of bridges, mean bridge area, fractal dimension of steatosis, luminosity of the parenchyma staining in the green component, number or fragments, and mathematical combination thereof; b. mathematically combining said at least two measured descriptors from step (a) in to a score, wherein said mathematical combination is a binary logistic regression, a multiple linear regression or any multivariate statistical analysis; and c. determining an increased risk of liver complications, an increased risk of mortality, or both, based on the value of the score calculated at step (b).

12. The method according to claim 11, wherein the at least one image is a histological section image, or an image of the entire liver organ or liver tissue.

13. The method according to claim 11, wherein the at least one image is obtained by an optical technique selected from the group consisting of: microscopic physical imaging, second harmonic generation (SHG), multiphoton imaging, coherent anti-Stokes Raman scattering (CARS), two-photon excitation fluorescence (TPEF), diffuse optical imaging and event-related optical signal; or by a non-optical technique selected from the group consisting of: radiography, nuclear medicine, photoacoustic methods, and thermal methods.

14. The method according to claim 11, further comprising measuring at least one non-invasive test marker and combining the resulting at least one measure with the measures of the two descriptors obtained at step (a), wherein the at least one non-invasive test marker is selected from the group consisting of a biomarker, a clinical data, a physical data, and/or a score selected from Enhanced Liver Fibrosis test (ELF™) FibroSpect™, AST to Platelet Ratio Index (APRI), Fibrosis-4 (FIB-4), Hepascore, Fibrotest™, FibroMeter™, CirrhoMeter™, CombiMeter™, and InflaMeter™.

15. The method according to claim 11, further comprising measuring a binary non-invasive diagnostic test marker and combining the resulting measure with the measures of the two descriptors obtained at step (a).

16. The method according to claim 11, further comprising measuring a Metavir F score (FM score), an Ishak stage, or both, and combining the resulting measure with the measures of the two descriptors obtained at step (a).

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a combination of images showing examples of total fibrosis (A), mask of porto-septal fibrosis (B), porto-septal fibrosis (C) and peri-sinusoidal fibrosis (D) as measured by the method of the invention.

(2) FIG. 2 is a combination of images showing focuses at ×20 magnification of total fibrosis (A), mask of porto-septal fibrosis (B), porto-septal fibrosis (C) and peri-sinusoidal fibrosis (D) as measured by the method of the invention.

(3) FIG. 3 is a combination of images showing examples of the edge of a non straight liver specimen (A) and a straight liver specimen (B) as measured by the method of the invention.

(4) FIG. 4 is a combination of images showing examples of a liver biopsy (LB) mask (A), the skeleton with ramifications (B) and the main skeleton (C) representing its length as measured by the method of the invention.

(5) FIG. 5 is a combination of images showing examples of stellar fibrosis around a porto-septal region (A), bridges of fibrosis (B), nodule surrounded by fibrosis (C), the distance measured between the porto-septal regions (D) as measured by the method of the invention.

(6) FIG. 6 is a combination of images illustrating the concept of granularity. Examples of fragments (n=4) of the liver biopsy specimen at the beginning of the study of granularity (A), porto-septal fibrosis on the fragments (B), granules formed by breaking the fragments with porto-septal fibrosis (C) as measured by the method of the invention.

(7) FIG. 7 is a combination of images showing examples of small fragments detected for measuring the INDEX_FRAGMENTATION. The total number of fragments is the same for the 2 panels A and B (NB_FRAG=8). As the area of small fragments is more important in figure A the INDEX_FRAGMENTATION is higher (INDEX_FRAGMENTATION=67) than in figure B (INDEX_FRAGMENTATION=35).

(8) FIG. 8 is a bean plot of the distribution of the CSF prediction (on Y axis) as a function of METAVIR staging (X axis) in the 416 patients of the MALAH 1 population. The grey filled curve is a kernel density estimate. The small grey lines represent individual data values. The large horizontal grey line shows the chosen threshold and the dotted black line corresponds to the overall median.

(9) FIG. 9 is a graph showing the distribution of the logistic function for the F4 prediction (Y axis) as a function of patient values ranked in increasing value (X axis) in the 416 patients of the MALAH 1 population. The horizontal lines depict the different thresholds. Triangle dots correspond to the original METAVIR F4 stage values whereas the white circle dots correspond to the original METAVIR stages less than F4. The black points mark patient slides with significant discordance (a difference of two stages or more between original Metavir evaluation and its prediction).

(10) FIG. 10 is a graph showing the distribution of the METAVIR stage prediction by linear discriminant analysis (LDA; Y axis) as a function of the METAVIR stages (on X axis) in the 416 patients of the MALAH 1 population. The vertical lines are drawn according to the proportion of original METAVIR stage values. On the Y axis, points are drawn with respect to the predicted stage. Their order follows their position in the dataset.

(11) FIG. 11 is a histogram showing the agreement between the prediction of the Metavir (on Y axis) and the Metavir stages evaluated by an expert (on X axis) in the 416 patients of the MALAH 1 population. The heights of the boxes reflect the proportion of the prediction stages for the expert stage.

(12) FIG. 12 is a histogram showing the distribution of NASH diagnosis as a function of intervals 95% predictive values by NASH score.

(13) FIG. 13 is a combination of graphs related to the NASH score. (A) deciles of NASH score (X axis) plotted against the proportion of NASH diagnosis (Y axis). (B) box plots of NASH score as a function of 95% predictive values and NASH diagnosis.

(14) FIG. 14 is a combination of 4 TDM images of the liver, corresponding to 4 consecutive slices (A, B, C and D) spaced 10 mm.

(15) FIG. 15 is a combination of illustrations showing a slice of TDM image (A), the mask of the liver and spleen (B), and the edges of the liver and spleen (C).

(16) FIG. 16 is a combination of illustrations showing a slice of TDM image (A) and a zoom on the irregular edges with nodules (B) (see arrows).

(17) FIG. 17 is a combination of illustrations showing a thresholded slice of TDM image with a box surrounding the liver (A), the mask of the box (B) and the black pixels corresponding to the fat in the box (C).

(18) FIG. 18 is a combination of illustrations showing examples of angles detected for the measure of the edges angularity in the liver of a healthy patient (F0) or in a cirrhotic liver (F4).

(19) FIG. 19 is a combination of pictures illustrating the measurement of internal morphology descriptors. (A): Example of different lengths measured in an enlarged liver segment I in cirrhosis. D1 is the anterior-posterior length between the most anterior limit/edge of segment I and the most posterior limit/edge of segment I tangential to the left edge inferior vena cava. D2 is the transversal length between D1 (perpendicular to D1) and the most distant left limit/edge of segment I. D3 is the anterior-posterior length between the most anterior limit of segment I and the most anterior limit of inferior vena cava. D4 is the transversal length between D3 (perpendicular to D3) and the most distant left limit/edge of segment I. (B): Example of the surface measured in an enlarged liver segment I in cirrhosis. (C) Example of the furrow thickness (D_F1) and surface (Surf_F1) measured in cirrhosis. Furrow surface can be obtained by contouring limits either manually or by an automated process or both methods (semi-automated process). Furrow thickness can be the maximum thickness or the mean thickness.

(20) FIG. 20 is an illustration showing the edge smoothing liver used for measuring the ratio between the 2 perimeters and for evaluating the indentation of the liver.

(21) FIG. 21 is a combination of two graphs related to the APV3 score. (A) Correlation of APV3 score with E-FibroMeter2G score (EFM2G) as a function of Metavir F stages. (B) box plots (median quartiles, extremes) of APV3 score are better distributed as a function of Metavir F stages than those of the E-FibroMeter2G score (EFM2G).

EXAMPLES

(22) The present invention is further illustrated by the following examples.

Example 1: Automatic Measurement of Lesions in a Histological Image

(23) Methods

(24) Study Design

(25) The morphometric diagnosis by automated morphometry was designed to diagnose clinically significant fibrosis (CSF), cirrhosis and Metavir F stages. The diagnostic models were developed in the derivation population. These results were validated by using two kinds of population: Populations with reference Metavir staging by expert to validate the diagnostic accuracy of diagnostic models based on automated morphometry. Population with Metavir staging performed by a first line pathologist to compare the diagnosis made by automated morphometry with that performed in real life conditions.
The conditions of liver biopsy specimens were the same in all populations. Especially, specimen lengths were the same and staining and morphometry performed by the same engineer were made centrally in the HIFIH laboratory.
Populations
Derivation Population
416 patients with chronic viral hepatitis C and a length of digitized biopsy≧20 mm for stages F0 to F3 and ≧12.1 mm for F4 were included (18 F0, 169 F1, 116 F2, 59 F3, 54 F4).
Validation Populations

(26) Fibrosys population included 54 patients with chronic viral hepatitis C. Liver biopsy was performed at week 0 and repeated at week 96. These two subsets were considered as different populations to avoid data redundancies. Liver interpretation was made by a single expert.

(27) Vindiag 10 population included 83 patients with chronic viral hepatitis C having a baseline liver biopsy. Liver interpretation was made by a single expert.

(28) Fibrostar population included 285 patients with chronic viral hepatitis C having a baseline liver biopsy. Liver interpretation was made by different pathologists with various expertise corresponding to the initial diagnosis as performed in clinical practice.

(29) Liver Specimen

(30) All LB were taken by needle through transcostal route, formalin-fixed, included in paraffin blocks and stained with picro-sirius red. This staining highlights the fibrosis in red, healthy tissue in yellow/orange and steatosis appears in white (i.e. optically empty). We aimed at measuring both fibrosis and steatosis. The specimen slides were then fully scanned with an Aperio digital slide scanner (Scanscope CS System, Aperio Technologies, Vista Calif. 92081, USA) image processor that provided high quality 30,000×30,000 pixel images at a resolution of 0.5 μm/pixel (magnification×20). In order to facilitate sharing and storage of these images, they were compressed using JPEG2000 software with a quality Q=70. Thus, a slide with 16,000×22,000 pixels was decreased from 1 GB (before compression) to 30 MB.

(31) Each virtual slide was analyzed to obtain a variety of parameters describing lesions such as area of fibrosis, fractal dimension . . . . The combination of all these morphometric data allowed us to automatically diagnose CSF, F4 and METAVIR stage. Then, they were compared to those evaluated by the pathologist.

(32) Classical Morphometry

(33) A mask of the biopsy is automatically created (MASK.sub.LB) after detecting the white background of the image so that it is not taken into account in the analysis. An operator manually cleans the slide to remove the various artefacts present such as folds, dust, blood vessels, biliary tracts . . . . All these artefacts are removed from MASK.sub.LB.

(34) Area and Fractal Dimension of Fibrosis and Steatosis

(35) First, a fuzzy generalized classification (Ménard et al, Pttern recognition, 2002) process allows the merging of pixel intensities into three classes (fibrosis, healthy tissue, white areas) using the minimization of an original energy function. S.sub.FIB is the threshold of the fibrosis class and S.sub.W the threshold of the white class. Then, a specifically-developed expert system (Roullier et al, Conf Proc IEEE Eng Med Biol Soc, 2007:5575-5578) was applied on the previously obtained white-labelled areas to extract steatosis vacuoles and to eliminate blood vessels and biliary tracts (that had characteristics close to those of macro vesicles of steatosis).

(36) The expert system rules are based on the size of the regions (very small areas, considered as noise, are eliminated), the neighborhood regions (vessels surrounded by fibrosis are eliminated), the circularity of the region (4π*area/perimeter.sup.2) (non-round regions, e.g., biliary tracts, were eliminated), the Hough transform of the region (allowing the detection and retention of vesicle aggregates), and a statistical texture parameter (heterogeneous regions, e.g. blood vessels, were eliminated). The measure of the area of fibrosis (AOF) or steatosis (AOS) is equal to the ratio of pixels of fibrosis or steatosis divided by the number of pixels in the studied area:
AOF=Pix.sub.FIB/Pix.sub.MASK.sub._.sub.LB*100
AOS=Pix.sub.STEA/Pix.sub.MASK.sub._.sub.LB*100
where Pix.sub.FIB is the number of fibrosis pixels; Pix.sub.STEA is the number of steatosis pixels; and Pix.sub.MASK.sub._.sub.LB is the number of pixels of the study area.

(37) The “box-counting” method (Moal et al, Hepatology, 2002, 36:840-849) has been extensively used for measuring the fractal dimension of many histologic objects as a complexity index. The box-counting method provides the fractal dimension of Kolmogorov (D). The technique has been reported in details for biological structures.

(38) Briefly, a grid of square boxes (with ε pixels as the side length) resembling a chessboard was superimposed over the histologic image of threshold fibrosis. Boxes intersecting with collagen fibers were counted. Another chessboard grid was then used to cover the entire surface of the microscopic field. Thus, the total number (N) of boxes of sides (ε) required to completely cover the collagen fibers reflects the perimeter examined with the scale ratio ε. This step was repeated with ε varying until a size of 14 pixels, and data were plotted on a log-log graph (i.e., log [N] against log [ε]). The relationship between points was measured by linear regression analysis using the least square method; the slope D of the regression line corresponds to the fractal dimension D. We did the same measurement for the fractal dimension of steatosis by using the images showing steatosis. We called D.sub.F and D.sub.S the fractal dimension of fibrosis and steatosis, respectively.

(39) Area and Fractal Dimension of Porto-Septal and Peri-Sinusoidal Fibrosis

(40) The AOF measured throughout the LB specimen does not distinguish the two kinds of fibrosis which are observed on the slide (FIGS. 1 and 2). Fibrosis can start in the centro-lobular region (alcoholic liver disease) and/or in the periphery of the lobule or around the portal tracts (viral liver disease). This fibrosis grows gradually to form bands of fibrosis, or septa, called septal fibrosis. This septal fibrosis is extended between the portal tracts or between the portal tracts and the centro-lobular region. The cirrhosis stage occurs when these different septa entirely surround hepatocytes. When pathologists are assessing the Metavir stage, their judgment is based on the porto-septal fibrosis. The other form of fibrosis is located between the rows of hepatocytes and sinusoids, the equivalent of the capillary in the liver, i.e. in the intercellular space where there are many liver metabolic exchanges. This is called the pen-sinusoidal fibrosis. Although it is not taken into account in the fibrosis scores in clinical use, this fibrosis is important because it has a key role in the genesis of liver failure and portal hypertension.

(41) Consequently, we developed a mask detection of porto-septal regions (MASK.sub.PORT, FIGS. 1B and 2B) which allows us to distinguish these two types of fibrosis (porto-septal and pen-sinusoidal fibrosis). The creation of this mask requires a large number of morphometric treatments (regarding erosions, dilations, the size of fibrosis areas . . . ) which involves a long computing time and requires a lot of memory to process the data. This is the reason why we reduced the image dimensions. IM.sub.COL is the color image of the LB specimen resized with a scale factor R.sub.SCALE=4. IM.sub.GREEN (the green component of IM.sub.COL) is thresholded by S.sub.FIB in order to get a binary image of fibrosis (IM.sub.FIB). Peri-sinusoidal fibrosis lies between the rows of hepatocytes; therefore, it is sufficient to eliminate it by detecting the hepatocytes which have the intensity of healthy tissue. Concerning veins and vessels, we determined that if their diameter was less than 200 μm and if the fibrosis around was small compared to their size, they are not considered in the MASK.sub.PORT. Dilatations and erosions are applied on IM.sub.FIB in order to obtain compacted regions to study. Porto-septal regions have a lot of fibrosis and therefore the elements of fibrosis with large areas are considered in the MASK.sub.PORT. The fractal dimension of fibrosis into these regions is also a criterion to determine if it is in MASK.sub.PORT. Thin regions (small ratio between the perimeter of the area and its surface) with nodes (interconnection points in fibrosis filaments) are a sign of branched forms as observed in pen-sinusoidal fibrosis and are thus not considered as porto-septal fibrosis.

(42) Finally, we obtained MASK.sub.PORT that separate pen-sinusoidal fibrosis from porto-septal fibrosis. The lobular region is the region of the LB without porto-septal fibrosis. We measure the area and the fractal dimension of porto-septal fibrosis (AO_FPORT and DF_FPORT), the area and fractal dimension of peri-sinusoidal fibrosis (AO_FPS and DF_FPS), the area of lobular peri-sinusoidal fibrosis (AO_FPS_LOB) and the ratio of peri-sinusoidal fibrosis among the whole fibrosis (RATIO_FPS) as follows:
AO_FPORT=Pix.sub.PORT/PixMASK_LB*100
AO_FPS=Pix.sub.FPS/PixMASK_LB*100
AO_FPS_LOB=Pix.sub.FPS/(PixMASK_LB−Pix_Mask_Port)*100
RATIO_FPS=Pix.sub.FPS/Pix.sub.FIB.sub._.sub.TOT*100
where Pix.sub.PORT is the number of pixels of porto-septal fibrosis; Pix.sub.FPS is the number of pixels of peri-sinusoidal fibrosis; PixMASK_LB is the number of pixels of the total area of the LB specimen; Pix_Mask_Port is the number of pixels of the porto-septal area; and Pix.sub.FIB.sub._.sub.TOT is the number of pixels representing all the fibrosis (porto-septal and peri-sinusoidal fibrosis).

(43) Table 1 resumes all these classical morphometric measurements.

(44) TABLE-US-00002 TABLE 1 Classical lesions measured by personal automated morphometry (n = 10). Descriptors Lesions Lesion measurements (abbreviation) Fibrosis Area of fibrosis and fractal AOF/D.sub.F dimension of whole fibrosis Area of fibrosis and fractal AO_FPORT/DF_FPORT dimension of portal fibrosis Area of fibrosis and fractal AO_FPS/DF_FPS dimension of sinusoidal fibrosis Area of lobular sinusoidal AO_FPS_LOB fibrosis Ratio of portal and sinusoidal RATIO_FPS fibrosis Steatosis Area of steatosis and fractal AOS/D.sub.S dimension of steatosis
New Morphometry
Measure of Edge Linearity (DF_EDGE, PCT_RECT)

(45) The analysis of the edges of the LB specimen (FIG. 3) is an important descriptors to consider for Metavir staging. Indeed, the shape of the LB specimen changes according to the Metavir stages, the edges are very straight in low stages (F0, F1, F2) whereas they tend to become more and more curved and irregular in the high stages (F3, F4). That is why we measured automatically the fractal dimension (DF_EDGE) and the linearity percentage (PCT_RECT) of the edges on the LB specimen. At first, we detect the edges of the fragments on the LB specimen and, thanks to this mask (MASK.sub.EDGE), we combine two methods to optimize the detection of straight edges. Method 1 consists to apply the Hough transform to detect straight lines on MASK.sub.EDGE. The measurement of the Hough transform is well known in the domain of image processing to detect shapes (Duda et al, Comm ACM, 1972, 15:11-15). This gives a mask called MASK.sub.HOUGH containing only the edges of the mask detected as straight by the Hough transform. Method 2 consists to create a straight mask MASK.sub.RECT from the edge mask (MASK.sub.EDGE). For this, we first detect the corners with a Harris detector (Harris et al, Proceedings of the 4.sup.th Alvey Vision Conference, 1988:147-151) and then we keep the edge points separated by a sampling step=2.4 mm. The lines between all these points are drawn and we finally obtained a theoretical straight mask. MASK.sub.RECT represents the edges of MASK.sub.EDGE that are in common with this theoretical straight mask.

(46) Thanks to a combination of the 2 masks, we obtained a MASK.sub.RECTCOMB (MASK.sub.RECTCOMB=MASK.sub.HOUGH+MASK.sub.RECT). MASK.sub.RECTCOMB thus contains all the straightest edges of the LB specimen. They determine the following formula:
PCT_RECT=Pix.sub.MaskRectComb/Pix.sub.MaskEdge*100.
where Pix.sub.MaskRectComb is the number of pixels of MASK.sub.RECTCOMB; and Pix MaskEdge is the number of pixels of MASK.sub.EDGE.
DF_EDGE=Fractal dimension of the edge of the LB on MASK.sub.EDGE.
Measure of the LB Specimen Length (LB_LENGTH)

(47) The pathologists consider that the LB specimen length must be higher than 15 or 20 mm to be representative. Usually, it is manually measured directly on the slide but this is not very precise because the LB specimen is not always straight. So, we decided to measure it automatically (LB_LENGTH) on the virtual slide. We use the mask obtained after the elimination of the artefacts (MASK.sub.LB) and we apply a morphometric operation (skeletonization) which gives us a skeleton of the LB. Then the small ramifications are removed in order to keep the main skeleton (MASK.sub.SKELETON) that represents the LB's length (FIG. 4).
LB_LENGTH=Pix.sub.LB.sub._.sub.SKELETON*IM.sub.Resolution*R.sub.SCALE.
wherein Pix.sub.LB.sub._.sub.SKELETON is the number of pixels that represent the LB specimen length on MASK.sub.LB, IM.sub.Resolution is the resolution of the scanned image (0.5 μm) and R.sub.SCALE=4 is the scale factor used to resize the image.
Measure of LB_PERIMETER and LB_AREA

(48) We measure automatically the perimeter and the area of the LB (LB_PERIMETER, LB_AREA). We detect the edges of the fragments on the LB specimen and, thanks to this mask (MASK.sub.EDGE), the perimeter is calculated as follow:
LB_PERIMETER=Pix.sub.MaskEdge*IM.sub.Resolution*R.sub.SCALE.
LB_AREA=Pix.sub.MASK.sub._.sub.LB*IM.sub.Resolution.sub.2*R.sub.SCALE.sub.2.
wherein Pix.sub.MaskEdge is the number of pixels of MASK.sub.EDGE, Pix.sub.MASK.sub._.sub.LB is the number of pixels of the total area of the LB specimen, IM.sub.Resolution is the resolution of the scanned image (0.5 μm) and R.sub.SCALE=4 is the scale factor used to resize the image.
Measure of Stellar Fibrosis (AOF_STELLAR_TOT, AOF_STELLAR_EP, AOF_STELLAR_LOB, MEAN_STELLAR_PORT and MEAN_AO_PORT)

(49) The main feature to differentiate stage of fibrosis F0 from stage F1 and higher stages is the presence of star-branched fibrosis around porto-septal regions (FIG. 5A). Indeed, fibrosis that we called stellar fibrosis starts to appear at the stage F1 around the portal tracts. These are small and fine fibrils around a more dense area. Therefore, it is interesting to automatically quantify this stellar fibrosis among the total surface of the LB specimen (AOF_STELLAR_TOT), among the surface of porto-septal regions (AOF_STELLAR_EP) and among the surface of lobular regions (AOF_STELLAR_LOB). We also used the number of porto-septal regions (NB_PORT) to measure the mean area of stellar fibrosis (MEAN_STELLAR_PORT) and the mean area of porto-septal regions (MEAN_AO_PORT).

(50) To make this measurement, we detect fibrosis in a distance of 100 μm around the porto-septal regions of MASK.sub.PORT. In order to differentiate stellar fibrosis from peri-sinusoidal fibrosis and concentrated fibrosis in the portal tracts, we combine several morphometric operations (successive erosions and dilations) which enable to only keep the thin fibrils of fibrosis connected to the porto-septal regions. Stellar fibrosis is measured with the following formulas:
AOF_STELLAR_TOT=Pix_Fib_Stellar/PixMASK_LB*100
AOF_STELLAR_EP=Pix_Fib_Stellar/Pix_Mask_Port*100
AOF_STELLAR_LOB=Pix_Fib_Stellar/(PixMASK_LB-Pix_Mask_Port)*100
MEAN_STELLAR_PORT=Pix_Fib_Stellar/NB_PORT
MEAN_AO_PORT=Pix_Mask_Port/NB_PORT
where Pix_Fib_Stellar is the number of pixels detected as stellar fibrosis; PixMASK_LB is the number of pixels of the total area on the LB specimen; Pix_Mask_Port is the number of pixels of the porto-septal area in MASK.sub.PORT; and NB_PORT is the number of porto-septal regions in MASK.sub.PORT.
Measure of Bridging Fibrosis (NB_BRIDGE, RATIO_BRIDGE, AOF_BRIDGE, MEAN_SURF_BRIDGE, MEAN_PERIM_BRIDGE, MEAN_THICK_BRIDGE)

(51) Bridging fibrosis fibers occur at the Metavir stage F2. We use the term bridges when two portal tracts are interconnected (FIG. 5B). A bridge is defined as a structure lying between two thick elements. In general, pathologists consider that if there is a ratio of more than 50% of bridges, the Metavir stage will be higher than F2. It is therefore important to measure the number of bridges (NB_BRIDGE), their ratio among the porto-septal areas (RATIO_BRIDGE), the area of fibrosis in the bridges (AOF_BRIDGE) and the bridges thickness (MEAN_THICK_BRIDGE) which is obtained thanks to their perimeter (MEAN_PERIM_BRIDGE) and their surface (MEAN_SURF_BRIDGE). For each porto-septal region in MASK.sub.PORT, we apply morphometric operations such as high erosion followed by a small dilation. The aim is to observe a separation of the structures which determines the presence of bridges. At the end of the morphometric operations, a studied area with at least two elements is considered as a bridge and added in the MASK.sub.BRIDGE. We obtain the following parameters:
NB_BRIDGE is the number of bridges in MASK.sub.BRIDGE
RATIO_BRIDGE=Pix_Mask_Bridge/Pix_Mask_Port*100
AOF_BRIDGE=Pix_Fib_Bridge/PixMASK_LB*100
where Pix_Mask_Port is the number of pixels in MASK.sub.PORT; Pix_Mask_Bridge is the number of pixels in MASK.sub.BRIDGE; Pix_Fib_Bridge is the number pixels of fibrosis in the bridges; and PixMASK_LB is the number of pixels of the total area on the LB specimen.
MEAN_THICK_BRIDGE=MEAN_SURF_BRIDGE/MEAN_PERIM_BRIDGE*100
where MEAN_SURF_BRIDGE is the mean of pixels representing the surface of bridges and MEAN_PERIM_BRIDGE is the mean of pixels representing the perimeter of bridges.
Measure of Granularity (PCT_GRANULARITY)

(52) The architecture of liver fibrosis is modified in high METAVIR stages. Indeed, the fibrous bridges can disorganize the LB specimen (FIG. 6). The measure of the granularity percentage (PCT_GRANULARITY) aims at quantifying this destructuration of the LB specimen by porto-septal tracts.

(53) First, we count the number of fragments (NB_FRAG) in the mask obtained after the elimination of the artefacts (MASK.sub.LB). Then, we use the porto-septal mask (MASK.sub.PORT) in which we applied several dilatations in order to extend the porto-septal areas. We subtract MASK.sub.PORT from MASK.sub.LB in order to only observe the granules formed by breaking the fragments. Nb_Granules is the number of granules. PCT_GRANULARITY is the ratio between the number of fragments without destructuration and the number of granules obtained in these fragments after destructuration by porto-septal areas:
PCT_GRANULARITY=100−(NB_FRAG/Nb_Granules*100)
Measure of Fragmentation (INDEX_FRAGMENTATION)

(54) A LB specimen is sometimes fragmented. It depends on the METAVIR stage, especially for the F4 stage where the LB specimen could contain several small fragments. So, the measure of the fragmentation index can be useful for the high fibrosis stages with a little PCT_GRANULARITY because of a numerous fragmentation.

(55) As for measuring PCT_GRANULARITY, we use MASK.sub.LB. We detect on this mask the small fragments (FIG. 7) in order to obtain MASK.sub.FRAG.sub._.sub.SMALL. We consider as small a fragment with a surface under 2 mm.sup.2 or a bigger fragment with a surface under 3 mm.sup.2 but with a circularity up to 0.7. The fragmentation index is the ratio between the surface of small fragments detected and the total surface of the LB specimen:
INDEX_FRAGMENTATION=Pix_Mask_Frag_Small/Pix_Mask_Frag*100
where Pix_Mask_Frag is the number of pixels in MASK.sub.LB and Pix_Mask_Frag_Small is the number of pixel in MASK.sub.FRAG.sub._.sub.SMALL.
Measure of Nodules (PCT_NOD and NB_NOD)

(56) The stage F4, also known as cirrhosis, is mainly characterized by the formation of nodules (FIG. 5C). These nodules are the result of the disruption of fibrosis that circles regions of hepatocyte tissue (parenchyma). So, a nodule is a circular and non-fibrotic (without fibrous septa inside) area surrounded by fibrosis.

(57) The process is the same as the measure of the PCT_GRANULARY applying MASK.sub.PORT on MASK.sub.LB to study the granules obtained. We only keep the granules if they are circular. The nodules are not always perfectly round, that is why we choose a circularity threshold of 0.45. Among these round granules, we keep those that have at least 30% of fibrosis around (30% of the external border). We finally obtained a mask with regions tending towards nodularity (MASK.sub.NOD). Nodules were considered as definitive when PCT_GRANULARY was ≧80%. We use the following formulas:
PCT_NOD=mean of percentage of fibrosis around areas in MASK.sub.NOD.
NB_NOD=number of nodules in MASK.sub.NOD with more than 80% of fibrosis around.

(58) We also measured the number of nodules in MASK.sub.NOD with more than 30% of fibrosis around (NB_NOD_30). The process is the same as the measure of NB_NOD but this time we keep the nodules in MASK.sub.NOD with more than 30% of fibrosis around.
NB_NOD_30=number of nodules in MASK.sub.NOD with more than 30% of fibrosis around.
Measure of Portal Distance (DIST_EP_MEAN)

(59) As mentioned above, the location of portal tracts tends to be heterogeneous in the higher fibrosis stages. So they are no longer distributed regularly and are closer to each other. It seems therefore interesting to measure the average distance between these porto-septal regions (FIG. 5D). NB_FRAG is the number of fragments in MASK.sub.LB. Nb_EP.sub.n is the number of porto-septal regions present on the fragment n (n lies between 1 and NB_FRAG). For each n fragment, we measure the minimum distance Dmin.sub.n between all porto-septal regions present on MASK.sub.PORT. Dmoy.sub.n is the average distance between regions on the porto-septal fragment n.
Dmoy.sub.n=Dmin.sub.n/(Nb_EP.sub.n−1).
The average distance between porto-septal regions for all fragments is called DIST_EP_MEAN:
DIST_EP_MEAN=(Dmoy.sub.1+Dmoy.sub.2+ . . . Dmoy.sub.NB.sub._.sub.FRAG)/NB_FRAG.

(60) Table 2 resumes all these new morphometric measurements.

(61) TABLE-US-00003 TABLE 2 New lesions related to liver fibrosis measured by the method of the invention for the diagnosis of CSF, F4 and Metavir stages (n = 21). Lesions Descriptor (abbreviation) Directly related to fibrosis: Stellar fibrosis AOF_STELLAR_TOT AOF_STELLAR_EP AOF_STELLAR_LOB MEAN_STELLAR_PORT MEAN_AO_PORT NB_PORT Bridges NB_BRIDGE RATIO_BRIDGE AOF_BRIDGE MEAN_THICK_BRIDGE MEAN_SURF_BRIDGE MEAN_PERIM_BRIDGE Granularity PCT_GRANULARITY Nodules PCT_NOD NB_NOD NB_NOD_30 Portal distance DIST_EP_MEAN Indirectly related to fibrosis: Fragmentation INDEX_FRAGMENTATION NB_FRAG Edge linearity PCT_RECT DF_EDGE LB length LB_LENGTH LB perimeter LB_PERIMETER LB area LB_AREA
Quality of the Staining

(62) The performance of our measures depends on the quality of the staining (coloration) of the LB. Indeed, with a pale color of fibrosis, the detection could miss some porto-septal regions, and consequently we could underestimate the classification (CSF, F4 and Metavir). Usually, the pathologist excluded the cases with a poor coloration. That is why we decided to automatically detect the LB to be excluded due to poor coloration. All the measures of intensity luminosity are applied on the three components of the image (RGB: Red, Green and Blue). The luminosity intensity of the LB is calculated by averaging all this pixel intensity and we obtain a mean intensity for each component: ILbR (the mean intensity of the LB on the red component), ILbG and ILbB. We do the same for the intensity of fibrosis (IfibR, IfibG, and IfibB) and for the parenchyma (IparenchymaR, IparenchymaG, and IparenchymaB).

(63) The quality of the coloration is also bad if the fibrosis intensity is closed to the intensity of the parenchyma or the LB. Thus, there is a weak contrast between the fibrosis and the parenchyma or LB. We developed a measure of these contrast as follows:

(64) $\mathrm{Contrast\_Fib}\mathrm{\_Parenchyma}=\sqrt{{\left(\mathrm{IparenchymaR}-\mathrm{IfibR}\right)}^2+{\left(\mathrm{IparenchymaG}-\mathrm{IfibG}\right)}^2+{\left(\mathrm{IparenchymaB}-\mathrm{IfibB}\right)}^2}$$\mathrm{Contrast\_Fib}\mathrm{\_Lb}=\sqrt{{\left(\mathrm{ILbR}-\mathrm{IfibR}\right)}^2+{\left(\mathrm{ILbG}-\mathrm{IfibG}\right)}^2+{\left(\mathrm{ILbB}-\mathrm{IfibB}\right)}^2}$

(65) Table 3 resumes all these measurements of luminosity intensity.

(66) TABLE-US-00004 TABLE 3 New measurements of luminosity intensity describing the quality of the coloration of the LB specimen (n = 11). Luminosity characteristic Descriptor (abbreviation) Fibrosis luminosity IfibR IfibG IfibB Parenchyma luminosity IparenchymaR IparenchymaG IparenchymaB Overall luminosity (of specimen) ILbR ILbG ILbB Luminosity contrast between fibrosis Contrast_Fib_Parenchyma and parenchyma Luminosity contrast between fibrosis Contrast_Fib_Lb and specimen (overall luminosity)
Statistical Models

(67) We used a classical two-step modeling strategy: a building step on all the data and an external validation step using two other datasets.

(68) After a first quick univariate study of all variables on the whole set of original data (416 slides) and a second analysis for each of the five METAVIR stages, we decided to perform bivariate analysis on quantitative variables using Spearman's rank correlation coefficient (r.sub.s) because (i) the normality assumption was not often fulfilled and (ii) some subsample had a small number of slides. Since regression techniques work better when variables are not strongly correlated, we partitioned all the original variables into clusters that were composed of variables that were altogether highly correlated (r.sub.s>0.8) and we then chose the most relevant variables for each cluster, thus reducing the 31 original variables to a set of 25 variables that were not too correlated. For the two binary outcomes (CSF and cirrhosis) we looked for the best binary logistic regression (BLR) model whereas for the 5-valued ordinal outcome METAVIR staging we looked for the best discriminant model using linear, quadratic, flexible, mixture techniques and other methods of logit analysis for ordered categories. All analyses were realized using the R software (Team RDC, available from http://www.R-project.org).

(69) After a thorough comparison of the outputs of the numerous programs for selecting the best variables (using both forward selection, backward elimination, and best subset selection procedures), we were able to end up for each of the three diagnostic targets with a statistical model that can be considered the best of all models, taking into account the performance of the model and the minimality of the number of variables. It has to be noted that for the ordinal variable METAVIR staging, the best model was found using a linear discriminant analysis method (LDA).

(70) We then proceed with the description of the model and the study of the predicted values. For the BLR regressions, the thresholds were determined with respect to the prevalence of the target (using an a priori value for CSF and an a posteriori value for cirrhosis). After the study of the regression coefficients and their entropy/information criteria (BIC, Akaike . . . ), we computed the area under the receiver operator characteristic curve (AUROC) and its confidence interval, the confusion matrix and we performed the statistical analysis and plotting of the four groups (true/false positive/negative). Concordance indexes and classification rates were also computed and reported.

(71) For the LDA model, we tested several imputations methods to deal with the probabilities of each ordinal value including maximum probability value imputation, class weighted imputation, two-stage values such as Fi±1, Fx/y or FX/y (the bigger value indicating an increased preference for the stage) . . . before computing also concordance and classification rates using here weighted distances between the stages. Finally, we decided to keep the standard method maximum probability value imputation since it gives good results without the introduction of arbitrary thresholds.

(72) For each of the computed models, the misclassification error was introduced as a new binary value and two new LBR regressions were performed using this value as target (i) with all the original 25 retained variables, (ii) using only the selected descriptors of the model, in order to analyze the influence of the variables on the misclassification. Since we knew the original value of METAVIR stage for the slides, it was possible to analyze the severe misclassification errors (discordance of more than one stage).

(73) The performance of a model is well known to be overestimated when evaluated on the data that led to the model (optimism bias). So we checked the three models against two other populations. The first one, called here FIBROSYS, used slides of 54 patients at two distinct periods, week 0 (W0) and week 96 (W96). The second file of data for the external validation, VINDIAG10, was composed of 83 patient slides selected from a cohort respecting the same conditions as the original data (METAVIR stage F4 or a length of digitized biopsy≧20 mm). Table 4 shows the distribution of the METAVIR stages for all the datasets.

(74) TABLE-US-00005 TABLE 4 Distribution of the METAVIR stages for the four datasets (populations). METAVIR stages Population F0 F1 F2 F3 F4 Total MALAH 1 18 169 116 59 54 416 FIBROSYS W0 0 5 21 20 8 54 FIBROSYS W96 0 4 21 20 9 54 VINDIAG 10 2 25 18 9 29 83
Results
Derivation Population
CSF Prediction

(75) The best BLR model includes 5 descriptors among those described previously: AOF_STELLAR_TOT, MEAN_THICK_BRIDGE, NB_BRIDGE, PCT_NOD, and PCT_RECT. The threshold value, “a priori” set to 0.5, was kept because of the prevalence of CSF (55%) and the good specificity and sensitivity of the associated model. With this threshold, we correctly classified almost 87.3% of the patients and we obtain an AUROC of 0.96. In the discordance matrix, we have the following results: 165 true-negative (TN), 198 true-positive (TP), 22 false-positive (FP) and 31 false-negative (FN). The results obtained (FIG. 8) confirm that this is a very good model to predict CSF. Table 5 gives the overall results of performance for the original dataset (MALAH 1) and Table 5b provides the model coefficients.

(76) TABLE-US-00006 TABLE 5 Overall results for the CSF prediction model. Correct classification rate using an a priori 0.5 threshold value. LWR and UPR: lower and upper bounds of the 5% confidence interval of the AUROC value. Correctly Patients AUROC classified Population (n) Value LWR UPR (%) Discordance MALAH 1 416 0.957 0.940 0.973 87.3 0.127 FIBROSYS 54 0.857 0.724 0.990 81.5 0.185 W0 FIBROSYS 54 0.895 0.787 1.003 81.5 0.185 W96 VINDIAG 10 83 0.880 0.804 0.955 80.7 0.193

(77) TABLE-US-00007 TABLE 5b Coefficients of the binary logistic regression for CSF prediction. Std. # Variable Lower Coefficient Upper Error z-value Prob. constant −1.0116 1.3268 3.7255 1.2028 1.1031 0.2700 1 AOF_STELLAR_TOT 15.1400 22.0255 29.7866 3.7235 5.9153 0.0000 2 PCT_RECT −0.1398 −0.0996 −0.0630 0.0195 −5.1098 0.0000 3 MEAN_THICK_BRIDGE 0.0075 0.0189 0.0312 0.0060 3.1294 0.0018 4 NB_BRIDGE 0.1234 0.3480 0.5887 0.1182 2.9439 0.0032 5 PCT_NOD 0.0015 0.0159 0.0302 0.0073 2.1798 0.0293 Std. error: standard error: standard deviation. Prob;: probability Lower: lower limit of 95% confidence interval. Upper: upper limit of 95% confidence interval.
Cirrhosis Prediction

(78) The best BLR model includes 6 descriptors to predict cirrhosis (F4): PCT_RECT, PCT_NOD, AOF_STELLAR_EP, PCT_GRANULARITY, DIST_EP_MEAN, and INDEX_FRAGMENTATION. As the logistic function (FIG. 9) is not symmetric, there is no reason to use the a priori threshold of 0.5. Our rule was to maximize cirrhosis sensitivity since this an important clinical diagnosis. In this model, the predictive threshold is selected “a posteriori” at 0.1567 providing a specificity of 0.964 and a sensitivity of 1. With this threshold, we correctly classified 96.6% of the patients and we obtain an AUROC of 0.994, thus showing an excellent model. Table 6 gives the overall results of performance for the original dataset (MALAH 1) and Table 6b provides the model coefficients.

(79) TABLE-US-00008 TABLE 6 Overall results for the cirrhosis prediction model. Correct classification rate using an a posteriori 0.1567 threshold value. LWR and UPR: lower and upper bounds of the 5% confidence interval of the AUROC value. Correctly Patients AUROC classified Population (n) Value LWR UPR (%) Discordance MALAH 1 416 0.994 0.989 0.999 96.9 0.034 FIBROSYS 54 0.978 0.943 1.01 87.0 0.130 W0 FIBROSYS 54 0.946 0.885 1.006 85.2 0.148 W96 VINDIAG 10 83 0.968 0.928 1.01 91.6 0.084

(80) TABLE-US-00009 TABLE 6b Coefficients of the binary logistic regression for cirrhosis prediction. Std. # Variable Lower Coefficient Upper Error z-value Prob. constant −0.1613 5.7840 12.2506 3.1138 1.8576 0.0632 1 PCT_RECT −0.2318 −0.1304 −0.0438 0.0472 −2.7617 0.0058 2 PCT_NOD 0.0345 0.0808 0.1409 0.0271 2.9744 0.0029 3 AOF_STELLAR_EP −4.9658 −3.2450 −1.9505 0.7590 −4.2751 0.0000 4 DIST_EP_MEAN −11.5238 −6.0570 −1.6248 2.4783 −2.4440 0.0145 5 PCT_GRANULARITY 0.0461 0.0880 0.1432 0.0243 3.6275 0.0003 6 INDEX_FRAGMENTATION 0.0326 0.0732 0.1227 0.0226 3.2329 0.0012 Std. error: standard error: standard deviation. Prob.: probability Lower: lower limit of 95% confidence interval. Upper: upper limit of 95% confidence interval.
METAVIR Stage Prediction

(81) At the difference of the previously predictions (CSF and F4), we apply here a linear discriminant analysis (LDA) to predict the METAVIR stages. The final model includes 8 descriptors: PCT_GRANULARITY, PCT_RECT, PCT_NOD, RATIO_FPS, DF_FPS, DF_FPORT, RATIO_BRIDGE and INDEX_FRAGMENTATION. The LDA does not provide a unique METAVIR score but it indicates a belonging probability for each stage. We use the classical method to affect a class by selecting the highest probability since the other imputations either (i) use arbitrary probability values or (ii) show no statistically better performance. The average discordance is of 0.315 and there is no patient with significant discordance (with a difference of 2 stages or more between METAVIR stage and prediction according to the classical definition used with non-invasive tests of liver fibrosis). The results obtained (FIGS. 10 and 11) confirm that it is also a good model. Table 7 gives the overall results of performance for the original dataset (MALAH 1). Agreement between original Metavir stages and predicted stages was very good according to weighted kappa index=0.868 (Table 7b). Tables 7c and 7d provides the model characteristics. Table 7e shows no significant discordance (≧2 F) between original and predicted stages.

(82) TABLE-US-00010 TABLE 7 Overall results for the METAVIR stage prediction model. Discordance value using equal or proportional distances between METAVIR stages. Correct classification rate for all METAVIR or F4 stage(s). Correctly Patients Discordance classified (%) Population (n) Equal Proportional All F4 MALAH 1 416 0.315 0.315 68.5 75.9 FIBROSYS W0 54 0.407 0.407 59.3 75.0 FIBROSYS W96 54 0.463 0.500 53.7 66.7 VINDIAG 10 83 0.289 0.337 71.1 82.8

(83) TABLE-US-00011 TABLE 7b Overall agreement between the METAVIR stage and its prediction model. Kappa coefficient lower estimate upper Unweighted 0.493 0.555 0.618 Weighted 0.844 0.868 0.891

(84) TABLE-US-00012 TABLE 7c Prior probabilities used in the LDA model for the prediction of Metavir stages. F0 F1 F2 F3 F4 0.0433 0.4062 0.2788 0.1418 0.1298

(85) TABLE-US-00013 TABLE 7d Coefficients of the LDA model for the prediction of Metavir stages. # Variable LD1 LD2 LD3 LD4 1 PCT_GRANULARITY −0.0147 −0.0459 −0.0217 0.0406 2 PCT_RECT 0.0424 0.0039 0.0264 −0.0378 3 PCT_NOD −0.0115 0.0019 −0.0063 −0.0055 4 RATIO_FPS 0.0479 −0.0340 −0.0219 0.0150 5 DF_FPS −5.2464 8.9728 3.6717 1.7888 6 DF_FPORT 4.3518 −6.5818 −10.3962 −6.7348 7 RATIO_BRIDGE −0.0181 −0.0005 0.0047 −0.0344 8 INDEX_FRAGMENTA- −0.0107 −0.0244 0.0442 −0.0105 TION LDi is the coefficient of the i.sup.th discriminant function.

(86) TABLE-US-00014 TABLE 7e Distribution of original METAVIR stage and predicted stages. Metavir Prediction F0 F1 F2 F3 F4 Original F0 4 14 0 0 0 F1 6 136 27 0 0 F2 0 32 71 13 0 F3 0 0 17 32 9 F4 0 0 0 13 41
Validation Populations
Model Performance

(87) Diagnostic accuracy of the diagnostic models based on automated morphometry was validated in Fibrosys and Vindiag 10 populations. Table 5 gives the overall results for CSF diagnosis in the three external validation datasets. Table 6 gives the overall results for cirrhosis diagnosis in the three external validation datasets. Table 7 gives the overall results for the Metavir stages in the three external validation datasets. Globally, there was an expected decrease in accuracy due to the lack of optimism bias. Nevertheless, the accuracies were still very good.

(88) Application to Clinical Practice

(89) In the Fibrostar population, the reference was Metavir staging performed by the central expert pathologist of Angers. The agreement between reference Metavir staging and second measurement was better with diagnosis by automated morphometry than initial diagnosis by local first line pathologist (Table 8).

(90) TABLE-US-00015 TABLE 8 Agreement between reference diagnosis by central expert and diagnosis by automated morphometry method of the invention or local pathologist in the 285 patients of the Fibrostar population. For Metavir F stages we used three kinds of agreement index. As Fibrosis staging is an ordinal variable, weighted kappa or intra-class correlation coefficient should be preferred. Best values are in bold characters. Agreement of reference diagnosis with: F stages: Morphometry Local pathologist kappa 0.515 0.611 weighted kappa 0.881 0.865 intra-class correlation coefficient 0.934 0.929 CSF (kappa) 0.733 0.733 F4 (kappa) 0.900 0.827

Example 2: Reliability

(91) The accuracy of prediction is imperfect. These inaccurate results are not due to chance and can be predicted by statistical analysis providing significant independent predictors of accuracy. This defines the reliability analysis. Thus, one can calculate reliability classes where the accuracy varies and predictors can be different.

(92) For example, a perfect result has 100% accuracy. A reliability analysis can determine reliability classes from 0 to 10, 20, 30, 40, 50, 60, 70, 80, 90% accuracy with specific predictors for each reliability class.

(93) Predictors are issued from the list of descriptors of the invention.

(94) For example, we provide here different predictive models obtained in different populations by binary logistic regression.

(95) Tables of the Coefficients of the Binary Logistic Regression for Prediction of Accuracy for Clinically Significant Fibrosis

(96) TABLE-US-00016 Model # 1, AUROC = 0.656 Variable Lower Coefficient Upper Std. error z-value Prob (Intercept) 1.9388 4.2135 6.6344 1.1949 3.5263 0.0004 MEAN_THICK_BRIDGE −0.0181 −0.0097 −0.0014 0.0042 −2.3029 0.0213 PCT_RECT −0.0687 −0.0358 −0.0044 0.0164 −2.1903 0.0285 Std. error: standard error: standard deviation. Prob.: probability Lower: lower limit of 95% confidence interval. Upper: upper limit of 95% confidence interval.

(97) TABLE-US-00017 Model # 2, AUROC = 0.893 Variable Lower Coefficient Upper Std. error z-value Prob MEAN_AO_PORT 12.9522 47.5080 86.5260 18.6618 2.5457 0.0109 MEAN_PERIM_BRIDGE −0.0012 −0.0006 −0.0001 0.0003 −2.4062 0.0161 RATIO_BRIDGE −0.1333 −0.0714 −0.0106 0.0310 −2.3080 0.0210 PCT_NOD −0.0558 −0.0296 −0.0051 0.0128 −2.3041 0.0212 AOF_BRIDGE 1.1042 4.8274 9.3013 2.0965 2.3026 0.0213 DF_FPS −65.6751 −33.6352 −5.0076 15.2713 −2.2025 0.0276

(98) TABLE-US-00018 Model # 3, AUROC = 0.820 Variable Lower Coefficient Upper Std. error z-value Prob LB_PERIMETER 0.0210 0.0775 0.1366 0.0293 2.6474 0.0081 AOF_BRIDGE 0.5897 1.9735 3.5409 0.7519 2.6246 0.0087 DF_FPORT 1.3641 7.0186 12.7720 2.8979 2.4219 0.0154 RATIO_FPS 0.0200 0.1092 0.2009 0.0460 2.3749 0.0176 LENGTH −0.1538 −0.0813 −0.0093 0.0368 −2.2093 0.0272 DF_FPS −32.0752 −16.2661 −1.1053 7.8551 −2.0708 0.0384 Anfractuosity −2.3940 −1.1928 0.0008 0.5997 −1.9888 0.0467 Anfractuosity = ratio between the native perimeter and the smoothed perimeter of the liver
Tables of the Coefficients of the Binary Logistic Regression for Prediction of Accuracy for Cirrhosis

(99) TABLE-US-00019 Model # 1, AUROC = 0.916 Coef- Variable Lower ficient Upper Std. error z-value Prob PCT_NOD −0.0719 −0.0369 −0.0061 0.0165 −2.2366 0.0253

(100) TABLE-US-00020 Model # 2, AUROC = 0.959 Variable Lower Coefficient Upper Std. error z-value Prob PCT_NOD −0.0924 −0.0550 −0.0232 0.0174 −3.1520 0.0016 PCT_RECT 0.0522 0.1474 0.2587 0.0518 2.8445 0.0044 N LENGTH −0.2933 −0.1619 −0.0454 0.0618 −2.6220 0.0087 LB_LENGTH 0.2795 0.8925 1.6303 0.3404 2.6217 0.0087 DF_FPORT 4.2902 17.9175 33.1817 7.2281 2.4789 0.0132 AOF_STELLAR_EP 0.7316 2.1216 4.1031 0.8765 2.4206 0.0155 NB_BRIDGE −1.2683 −0.6658 −0.1301 0.2851 −2.3352 0.0195 DF_FPS −74.6557 −37.7834 −7.1016 16.8191 −2.2465 0.0247 LB_PERIMETER −0.3474 −0.1687 −0.0105 0.0844 −1.9988 0.0456 N_LENGTH: length of numeric specimen; LB_LENGTH: initial length of fixed specimen

(101) TABLE-US-00021 Model # 3, AUROC = 0.984 Coef- Variable Lower ficient Upper Std. error z-value Prob PCT_NOD −0.6229 −0.2149 −0.0796 0.1087 −1.9776 0.0480

Example 3: Diagnosis of NASH

(102) Non-alcoholic fatty liver disease (NAFLD) is a frequent pathology. It encompasses the liver lesions linked to metabolic syndrome. It evolves from pure steatosis to non-alcoholic steato-hepatitis (NASH) and liver fibrosis. NASH includes several lesions: steatosis, hepatocyte ballooning, and lobular inflammation (Sanyal Hepatology 2011). An international expert team has described a NAFLD activity score (NAS) (Kleiner Hepatology 2005).

(103) We have measured several lesions in a cohort of 235 patients with NAFLD. Optical microscopy included classical lesions (steatosis degree, NAS, NASH and fibrosis). Morphometry with image analysis included automatic measurement of area of steatosis (AOS) and fractal dimension of steatosis (DS) as well as area of fibrosis (AOF) and fractal dimension of fibrosis (D.sub.F). We also defined the relative Area of Steatosis (rAOS) as the area of steatosis in the non-fibrotic area.

(104) Whatever the fibrosis stage, NAS was globally well correlated (p<0.001) with area of steatosis (AOS) (r.sub.s=0.746), relative Area of Steatosis (rAOS) (r.sub.s=0.761), and fractal dimension of steatosis (DS) (r.sub.s=0.794).

(105) Mean fractal dimension of steatosis (DS) was much higher when NASH was present: 1.652±0.100 vs 1.395±0.126 in the absence of NASH (p=0.001).

(106) Consequently, NASH was well predicted by relative Area of Steatosis (rAOS) with AUROC=0.919 with diagnostic accuracy at 85.6% or by fractal dimension of steatosis (DS): AUROC=0.936, diagnostic accuracy: 87.6%.

(107) The best cut-off to diagnose NASH was relative Area of Steatosis (rAOS)=8.2% or fractal dimension of steatosis (DS)=1.595.

(108) By stepwise binary logistic regression, NASH was diagnosed by the combination—called the NASH score—of fractal dimension of steatosis (DS), area of fibrosis (AOF) and relative Area of Steatosis (rAOS) with AUROC=0.953 and diagnostic accuracy=87.6%. The proportion of patients with 95% predictive values for NASH was: rAOS: 44.6%, DS: 63.7%, NASH score: 75.5% (p<0.001 between each proportion by McNemar test).

(109) FIG. 12 shows the distribution of NASH diagnosis as a function of intervals of 95% predictive values by NASH score. The mean NASH scores of the patients were the following:

(110) TABLE-US-00022 Determination of the presence of Mean NASH NASH with the NASH score NASH score Negative No 0.04 Doubtful 0.11 Yes N/A Intedeterminate No 0.43 Doubtful 0.47 Yes 0.55 Positive No 0.88 Doubtful 0.93 Yes 0.97

(111) FIG. 13A shows the deciles of NASH score (X axis) plotted against the proportion of NASH diagnosis (Y axis). FIG. 13B shows the box plots of NASH score as a function of 95% predictive values and NASH diagnosis. As shown in FIGS. 13A and B, with the 95% negative predictive value cut-off, there is no missed definitive NASH whereas with the 95% positive predictive value cut-off, there are only 3% of patients without NASH.

(112) Moreover, we tested our descriptors in a new population of 137 patients with NAFLD. The histological definition of NASH was the most recent as described in a consensus paper (Sanyal Hepatology 2011). The NASH is defined as borderline or definitive diagnosis.

(113) Results were the followings according to the two possible NASH definitions:

(114) TABLE-US-00023 NASH diagnostic Patient number Correctly target (biopsy lentgh) AUROC classified (%) Borderline + 107 (≧20 mm) 0.921 84.1 definitive 137 (all)   .sup.  0.891 80.3 Definitive 107 (≧20 mm) 0.868 78.5 137 (all)   .sup.  0.845 75.2
Model for Definitive NASH

(115) TABLE-US-00024 Standard Variable Coefficient deviation Z value Probability Constant −13.32199153 3.463405925 −3.846500 1.198170e−04 Contrast_Fib_Parenchyma −0.05635018 0.022324528 −2.524137 1.159826e−02 Fractal dimension of steatosis 8.23338304 2.039900986 4.036168 5.433132e−05 Biopsy length 0.01743691 0.008354337 2.087169 3.687290e−02 AOF_STELLAR_TOT 14.95754878 3.578657393 4.179654 2.919533e−05
Model for Borderline+Definitive NASH

(116) TABLE-US-00025 Standard Variable Coefficient deviation Z value Probability Constant −17.43729686 4.05374381 −4.301529 1.696235e−05 Contrast_Fib_Parenchyma −0.09204421 0.03120528 −2.949636 3.181486e−03 Fractal dimension of steatosis 14.51277356 3.09510099 4.688950 2.746106e−06 AOF_STELLAR_EP −0.40530236 0.20406950 −1.986100 4.702226e−02 AOF_STELLAR_TOT 15.77945921 4.22038919 3.738864 1.848540e−04

Example 4: Automatic Measurement of Lesions in a Radiological Image

(117) We developed image analysis from tomodensitometry (TDM) to allow an automated diagnosis of cirrhosis and other fibrosis stages. The advantage of the TDM images is that, contrary to a liver biopsy, we can evaluate the entire liver. Consequently, we dispose of more representative information. The way to proceed is the same as the liver biopsy method, except that we are not analysing one slice but several slices at intervals of 10 mm. Thus, for one patient we obtain approximately 15-20 images of 512×512 pixels with grey levels (FIG. 14).

(118) Semiology

(119) We can separate the measurements of descriptors in different categories (see Table 9):

(120) TABLE-US-00026 TABLE 9 Category Measurements External morphology Liver Anfractuosity: Native perimeter of the liver (mean and total perimeter) Smoothed perimeter Ratio between the 2 perimeters Largest perimeter among all the slices Indentation Liver height Nodularity of the edges Angularity Fat ratio in the box including the liver Abdominal liver fat ratio Spleen Ratio between the spleen area and the liver area Total spleen perimeter Spleen height Nodularity of the edges Internal morphology Hypertrophy of liver segment I (2 antero-posterior lengths D1, D3 and 2 transversal lengths D2, D4: see figure). Hypertrophy/surface/volume of liver segment I can be also expressed as any product of 2 to 4 lengths D1 to D4 The volume can be calculated with surface and the height of liver segment I Lengths of the liver segment IV: minimal and maximal anterior-posterior or transversal length Surface of liver segment IV can be also expressed as any product of 2 lengths The volume can be calculated with surface and the height of liver segment IV Lengths surface/volume of other liver segment: as described for liver segment IV Whole liver or liver segment volume obtained by recon- struction Whole liver or liver segment surface can be also obtained by contouring liver or liver segment limits either manually or by an automated process or both methods (semi auto- mated process) Ratio between segment I and segment IV dimensions Furrow thickness (D_F1) and surface (Surf_F1) (e.g. Arantius ligament) Diameter of the portal vein Internal nodularity in the liver Structure Heterogeneity of the density intensity in the image between several regions of interest (ROI) in the liver Fractal organization of the hepatic parenchyma
External Morphology

(121) As for the LB specimen, the analysis of the edges of the liver in TDM is an important descriptors to consider for the cirrhosis diagnosis. The edges tend to become more and more curved and irregular in the high fibrosis stages (Metavir F3, F4). The edges of spleen and liver may be assessed on a mask of the spleen and liver (FIG. 15). Therefore, we measured the characteristic of anfractuosity with the native perimeter of the liver (mean and total perimeter), the smoothed perimeter of the liver (FIG. 20) and the ratio between the native and smoothed perimeters. We also measured the largest perimeter among all the slices. The indentation of the liver can be interesting to evaluate the variation of the edges around smoothed edges. We completed the description of the liver external morphology with its height, the fractal dimension of its edges (FIG. 18), the nodularity of its curved and irregular edges (FIG. 16), and the angularity representing the angles present on the edges. We thus obtained a large number of variables describing the liver. Finally, we designed a box containing the liver wherein we measured the liver fat ratio (black pixels vs other pixels, FIG. 17). We also calculated in the same manner the abdominal fat ratio (black pixels vs other pixels contained in the abdomen). We did the same evaluation for the spleen and we evaluated the spleen height, the total spleen perimeter, the ratio between the spleen area and the liver area and the nodularity of the spleen edges.

(122) Internal Morphology (FIG. 20)

(123) Other interesting parameters are the hypertrophy of liver segment I (one) which is observed in the high Metavir fibrosis stages, the width of the liver segment IV, and the ratio between segment I and segment IV dimensions. The hypertrophy is measured with different lengths (see FIG. 19) and their mathematical combination (ratio and products). We used the inferior cave vein as a landmark for these measures: we obtained 2 antero-posterior lengths (D1 and D3, FIG. 19A), 2 transversal lengths (D2 and D4, FIG. 19A) and the surface of the segment I (Surf_S1, FIG. 19B). The furrow thickness (D_FI) and surface (Surf_F1), especially the Arantius ligament or furrow which tends to increase in cirrhosis, were also evaluated (FIG. 19C), as the the internal nodularity in the liver. The diameter of the portal vein was also measured.

(124) Structure

(125) We measured the heterogeneity of the density intensity in the image between several regions of interest (ROI) in the liver. Fractal organization of the hepatic parenchyma was also studied.

(126) Calculation

(127) The diagnostic value of radiological descriptors is calculated as follows.

(128) The dependent variable is the diagnostic target among those of pathological liver lesions: Fibrosis as binary target like significant fibrosis or cirrhosis or staging like Metavir staging. Steatosis.

(129) The independent variables are selected among the list of available radiological descriptors among those known in the literature plus the new signs described in the present application.

(130) The dependent variable is predicted by the independent variables thanks to appropriate statistical multivariate analysis like binary logistic regression or discriminant analysis. The multivariate analysis provides a score including one to several independent variables.

Example

(131) Population: 30 patients with chronic hepatitis C, liver biopsy (Metavir staging) and TDM.

(132) The independent variable was furrow surface:

(133) TABLE-US-00027 Diagnostic target Significant fibrosis Cirrhosis AUROC 0.882 0.893 Correctly classified (%) 79.3 75.9 Score coefficients: Constant −1.148 −2.645 Furrow surface 2.793 2.026 P value for furrow 0.031 0.005 surface

Example 5: Automatic Measurement of Lesions in a Radiological Image

(134) List of Radiological Descriptors

(135) The 48 following liver descriptors (Table 10) were measured by semi-automated digitized morphometry on images provided by tomodensitometry:

(136) TABLE-US-00028 TABLE 10 Site Descriptor External morphology Liver 1. Mean native liver perimeter (MNLP) 2. Total native liver perimeter 3. Total smoothed liver perimeter 4. Ratio native/smoothed total liver perimeters (RNSTLP) 5. Largest liver perimeter (LLP) 6. Mean liver surface 7. Total liver surface 8. Ratio liver perimeter/liver surface (RLPS) 9. Liver indentation 10. Liver height 11. Fractal dimension of liver edges 12. Liver nodularity 13. Liver angularity 14. Liver fat ratio Abdomen 15. Abdominal fat ratio Spleen 16. Spleen height (SH) 17. Total spleen perimeter (TSP) 18. Mean total spleen perimeter (MTSP) 19. Total spleen surface 20. Mean total spleen surface 21. Ratio spleen/liver surfaces (RSLS) 22. Ratio spleen/liver perimeters 23. Nodularity of the spleen edges 24. Mean fractal dimension of spleen edges Internal morphology 25. Width of liver segment IV 26. Ratio between segment I and segment IV dimensions 27. Ratio segment I/liver surfaces (RS1LS) 28. First antero-posterior segment I length (D1) 29. First transversal segment I length (D2) 30. Second antero-posterior segment I length (D3) 31. Second transversal segment I length (D4) 32. Segment I surface 33. Arantius furrow thickness (AFT) 34. Arantius furrow surface 35. Internal nodularity 36. Portal vein diameter Liver structure 37. Mean total density 38. Standard deviation of total density (SDTD) 39. Coefficient of variation of total density 40. Median total density 41. Interquartile range of total density (IQRTD) 42. Ratio interquartile range/median of total density 43. Mean ROI density 44. Standard deviation of ROI density (SDROID) 45. Coefficient of variation of ROI density 46. Median ROI density 47. Interquartile range of ROI density 48. Ratio interquartile range/median of ROI density * Descriptors in bold are independent descriptors in the following diagnostic scores.

(137) Descriptors in the Table above may be combined by division or multiplication, e.g. liver segment I hypertrophy can be described by D1×D2×D3×D4. Some descriptors can be applied to other structures, e.g. dimensions of all furrows, spleen density . . . .

(138) Diagnostic Scores

(139) Development and evaluation of diagnostic scores was performed in a population of 107 patients with chronic liver disease of miscellaneous causes. All patients had liver tomodensitometry. The diagnostic reference was either liver biopsy in 91 patients or non-invasive fibrosis diagnosis by Elasto-FibroMeter V2G (E-FibroMeter.sup.2G) combining Fibroscan and FibroMeter V2G in 107 patients. Both references are based on fibrosis Metavir staging (F0 to F4). Scores were obtained by binary logistic regression with Metavir staging outcomes (targets) as diagnostic reference. Diagnostic targets were significant fibrosis (Metavir F≧2), severe fibrosis (Metavir F≧3) and cirrhosis (Metavir F=4).

(140) Liver Biopsy as Reference

(141) Scores were obtained by binary logistic regression with Metavir staging as diagnostic reference.

(142) The following Table 11 describes test composition and accuracy.

(143) TABLE-US-00029 TABLE 11 Diagnostic Variables targets Type Score for: List * TDM FS BT N AUROC Cirrhosis Pre 1 SDTD, D3, AFT, x 4 0.872 RSLS Pre 4 Pre 1 + TSP x 5 0.883 Fibroscan x 1 0.853 CirrhoMeter x 8 0.797 FibroMeter x 8 0.726 E-FibroMeter x x 8 0.866 Pre 5 Pre 4 + FS x x 6 0.922 Pre 6 Pre 4 + FS, GGT x x x 7 0.940 Pre 8 Pre 4 + FS, GGT, x x x 9 0.953 age, HA Significant fibrosis Pre 2 IQRTD, AFT, MNLP, x 5 0.910 MTSP, RSLS Fibroscan x 1 0.920 CirrhoMeter x 8 0.840 FibroMeter x 8 0.803 E-FibroMeter x x 8 0.937 Pre 9 Pre 2 + FS, PI, A2M x x x 7 0.982 Pre 10 Pre 2 + FS, A2M, x x x 8 0.985 AST, PI Severe fibrosis Pre 3 RS1LS, D4, RNSTLP x 3 0.821 Pre 1 x 4 0.822 Pre 4 x 5 0.823 Fibroscan x 1 0.898 CirrhoMeter x 8 0.795 FibroMeter x 0.705 E-FibroMeter x x 0.894 Pre 11 Pre 3 + FS x x 4 0.905 Pre 12 Pre 2 + FS, AST x x x 6 0.929 HA: hyaluronic acid, PI: prothrombin index, A2M: alpha2-macroglobulin, FS: Fibroscan, BT: blood tests, TDM: tomodensitometry, Pre: predictive score. *Scores and descriptors in bold characters include only descriptors provided by tomodensitometry.

(144) As shown Table 11, for cirrhosis diagnosis, tomodensitometry outperforms all other non-invasive tests. Moreover, for significant fibrosis diagnosis, tomodensitometry improves the best non-invasive test (E-FibroMeter).

(145) E-FibroMeter as Reference

(146) Scores were obtained by multiple linear regression with E-FibroMeter V2G as diagnostic reference.

(147) E-FibroMeter score was divided into binary diagnostic targets: significant fibrosis and cirrhosis according to respective maximum Youden cut-off.

(148) TABLE-US-00030 TABLE 12 Composition (Table 12) Descriptors N Type Scores List (abbreviation) * Descriptor TDM FS BT APV 1 AFT, MNLP, RSLS, SDTD, 5  x SH APV 2 Pre 9, Pre 2, Pre 5, SDROID, 6+ x x x RLPS, LLP APV 3 Pre 9, Pre 5, SDROID, Pre 1 4+ x x x APV: adjusted predictive value * Scores and descriptors in bold characters include only descriptors provided by tomodensitometry.

(149) TABLE-US-00031 TABLE 13 Correlations (Rs, Table 13) E-FibroMeter Metavir F N patients 99 82 APV 1 0.552 0.498 APV 2 0.883 0.723 APV 3 0.918 0.779 E-FibroMeter — 0.755 Fibroscan 0.926 0.718

(150) TABLE-US-00032 TABLE 14 Accuracy (AUROC, Table 14) AUROC Reference Metavir E-FibroMeter Diagnostic target Significant Significant Score * fibrosis Cirrhosis fibrosis Cirrhosis N patients 82 82 99 99 APV 1 0.794 0.777 0.714 0.788 APV 2 0.958 0.832 0.975 0.967 APV 3 0.973 0.877 0.947 0.977 E-FibroMeter 0.937 0.866 — — Fibroscan 0.920 0.853 0.918 0.973 Pre 2 0.910 0.805 0.746 0.795 Pre 4 0.796 0.883 0.740 0.810 Pre 10 0.985 0.900 0.957 0.991 * Scores in bold characters include only descriptors provided by tomodensitometry.

(151) As shown in Table 14 and in FIG. 21, for significant fibrosis diagnosis, tomodensitometry (APV 3 score) improves the accuracy of other non-invasive tests. Moreover, APV 3 score defined with E-FibroMeter outperforms Pre 2 score defined with Metavir staging.

(152) For cirrhosis diagnosis, tomodensitometry (APV 3 score) improves the accuracy of other non-invasive tests.

(153) In conclusion, several multivariate scores including some independent radiological descriptors, measured by semi-automated digitized morphometry, either alone (tomodensitometry) or combined with elastometry and/or blood markers have a high diagnostic accuracy for significant liver fibrosis and cirrhosis.

Example 6: Prognostic Score

(154) Methods

(155) The Vindiag 10 population included 204 patients with CHC (chronic Hepatitis C) and a baseline LB (Liver Biopsy). In this population, follow-up was 10.2±3.7 years for mortality and 8.5±4.5 years for liver complications. During follow-up, sustained virological response (SVR) to antiviral treatment was observed in 40.2% of patients, overall mortality was 15.2%, liver-related mortality was 5.4% and liver complications occurred in 16.7% of patients. Mortality was recorded according to a national registry. Survival was analyzed with the Cox model.

(156) Results

(157) Histological Data

(158) In this analysis, independent predictors included all classical optical analysis (Metavir F stage), classical descriptors (such as, for example, area of fibrosis) and new descriptors (such as, for example, mean bridge area) including new morphometric scores (measured by mathematical combination according to the method of the invention, such as, for example, cirrhosis score shown in above Table 6b). Most of the independent prognostic information was provided by new descriptors of the invention whatever the prognostic outcome (Table 15). Classical Metavir staging (F.sub.M) had either no significant role or borderline role.

(159) TABLE-US-00033 TABLE 15 Independent predictors among histological data by Cox model of clinical events in the Vindiag 10 population. First predictor figures indicate the step in forward stepwise analysis. Overall Liver related Liver mortality mortality complications Patients Variables .sup.a χ.sup.2 Variables .sup.a χ.sup.2 Variables .sup.a χ.sup.2 All Mean area of 24.3 Mean area of 148.6 Mean bridge area 70.8 porto-septal porto-septal Fractal fibrosis fibrosis dimension of Metavir F.sub.M Lobular area of steatosis stellar fibrosis Cirrhosis score Number of nodules Number of bridges .sup.a The figure indicates the step in forward stepwise binary logistic regression χ.sup.2: overall significance of the model. Morphometric descriptors of the invention are indicated in italics
Combined Data

(160) Cox model survival analysis in the Vindiag 10 cohort included demographic data, SVR, 6 blood fibrosis tests, Metavir staging (F.sub.M) and the 44 morphometric variables. Detailed results are presented in Table 16.

(161) TABLE-US-00034 TABLE 16 Independent predictors by Cox model of clinical events in the Vindiag 10 population. First predictor figures indicate the step in forward stepwise analysis. Overall Liver-related Liver mortality mortality complications Patients Variables .sup.a χ.sup.2 Variables .sup.a χ.sup.2 Variables .sup.a χ.sup.2 All CirrhoMeter.sup.2G 46.5 CirrhoMeter.sup.2G 30.9 CirrhoMeter.sup.2G 76.6 p = 0.015 p < 0.001 p < 0.001 SVR Area of Area of p = 0.033 steatosis steatosis Fibrotest p = 0.003 p < 0.001 p = 0.010 SVR Portal distance p = 0.028 p = 0.013 Fibrotest Edge fractal p = 0.001 dimension Hepascore p = 0.047 p = 0.039 Nodule number p = 0.039 Liver CirrhoMeter.sup.2G 24.7 CirrhoMeter.sup.2G 32.3 CirrhoMeter.sup.2G 59.7 specimen ≧20 p < 0.001 p = 0.007 p < 0.001 mm or F.sub.M4 F.sub.M4 score Area of Area of p = 0.003 steatosis steatosis p = 0.017 p = 0.001 F.sub.M4 score Intensity p = 0.010 specimen blue Intensity of p = 0.005 specimen-blue Edge fractal p = 0.028 dimension p = 0.016 Fragments number p = 0.008 Intensity of parenchyma green p = 0.052 .sup.a The figure indicates the step in forward stepwise binary logistic regression χ.sup.2: overall significance of the model. SVR: sustained virological response. Morphometric descriptors of the invention are indicated in italics

(162) Briefly, whatever the prognostic outcome (overall mortality, liver mortality, liver complications) or population (all 204 patients or 110 patients with liver specimen length≧20 mm or F.sub.M4), CirrhoMeter.sup.2G was always the first independent predictor. There was always at least one new morphometric descriptor of the invention among other predictors. Particularly, area of steatosis was the second predictor for liver events (mortality or complications). Interestingly, morphometric F.sub.M4 score was an independent predictor for both overall and liver-related mortality in patients with adequate specimen (length≧20 mm or F.sub.M4). It should be noted that Metavir F.sub.M had no independent role.