METHOD AND SYSTEM FOR IMAGE NORMALISATION

Abstract

The present invention relates to a method and system for the transformation of raw mammograms to normalised presentation and where the pixel values are independent of imaging conditions. The performed method includes: contrast enhancement, for improved visibility of the breast tissue composition, whereby a region of the breast is segmented and a contrast-stretching algorithm applied to the segmented region to preferably create an enhanced raw image or mammogram; local ‘maximum’ transform, whereby a 2-dimensional first filter is designed to extract the maximum pixel value from a region of interest (ROI) to preferably create a local maximum image or map; ratio map derivation, whereby the pixel value of the ratio map measures a relative response of the said pixel to its local maximum thus capturing the difference between breast composition regardless of mammogram variations.

Claims

1-39. (canceled)

40. A method for transformation of a raw mammogram including an image of a breast to normalise presentation, comprising the steps of: 1) creating a contrast enhanced mammogram for improved visibility of breast tissue composition, wherein a region of image of the breast in the raw mammogram is segmented and a contrast-stretching algorithm applied to the segmented region; 2) creating a local maximum map of the contrast enhanced mammogram by performing a local maximum transform, wherein a 2-dimensional first filter extracts a maximum pixel value from a region of interest (ROI); and 3) deriving a ratio map which includes a pixel-wise ratio between the contrast enhanced mammogram and the local maximum map, wherein a pixel value of the ratio map is a measure of a relative response of the pixel of the ratio map to its local maximum in the raw mammogram.

41. A method according to claim 40 wherein in step 1) the breast region is segmented as follows, from the raw mammogram, a binary image is derived where non-zero pixels indicate the location of the breast in the corresponding raw mammogram image.

42. A method according to claim 40 wherein applying the contrast-stretching algorithm includes segmenting the breast region in the raw mammogram and scaling pixel values in the breast region.

43. A method according to claim 40 wherein in step 1) pixel values inside the breast region in the raw mammogram are scaled to a range between 0 and 1 followed by a logarithm transform.

44. A method according to claim 40 wherein the contrast stretching algorithm includes a correction performed on scaled pixel values of the raw mammogram.

45. A method according to claim 40 wherein the contrast stretching algorithm includes a correction Performed on scaled pixel values of an image in which the breast region has saturated scaled pixel values.

46. A method according to claim 40 wherein in step 1) image saturation thresholds are determined such that guided by the segmentation, scaled pixels in the breast region of the raw mammogram are sorted in ascending order based on pixel values.

47. A method according to claim 46 in which the correction includes saturation parameters which quantify low and high saturation thresholds of pixels in the breast region.

48. A method according to claim 41 wherein a top and bottom percentile of the pixel values in the breast region are determined as saturation thresholds.

49. A method according to claim 48 in which the low and high saturation thresholds are in a range of 0.5% to 5% of the lowest and highest respectively of the pixels according to their pixel values.

50. A method according to claim 41 wherein in the raw mammogram, pixel values smaller than a low saturation threshold are set to the low saturation threshold, and pixel values greater than a high saturation threshold are set to the high saturation threshold to make a saturated image.

51. A method according to claim 50 wherein the saturated image has a pixel value range between low and high saturation thresholds.

52. A method according to claim 41 wherein the contrast-stretching algorithm includes gamma correction on a saturated raw mammogram where saturation parameters are determined from scaled pixels in the breast region.

53. A method according to claim 46, wherein gamma correction maps the pixel values of a saturated image to a nonlinear range between specific bottom and top output thresholds.

54. A method according to claim 46 wherein the bottom and top output thresholds are set to parameters that yield a higher contrast between fatty and dense, i.e. (fibroglandular), tissue than in the raw mammogram.

55. A method according to claim 40 wherein the first filter extracts the maximum pixel value from a 3×3 image region and replaces the center pixel value with the regional maximum.

56. A method according to claim 40 wherein a customised second filter is applied to the ratio map to generate a BACs probability image.

57. A method according to claim 56 wherein the customised second filter uses Frangi's vessel enhancement filter, tuned to accommodate the characteristics of BACs and a mammogram, to obtain a measure of vessel-probability.

58. A method according to claim 40 wherein as pixel values of the ratio map range from 0 to 1, a rescaling to an 8, 16, 32- or 64-bit integer image obtains a normalised range.

59. A system for transformation of a raw mammogram including an image of a breast to normalise presentation, the system comprising a processor in communication with a memory arranged to: 1) create a contrast enhanced mammogram for improved visibility of breast tissue composition, wherein a region of image of the breast in the raw mammogram stored in the memory is segmented and the processor is arranged to implement a contrast-stretching algorithm applied to the segmented region; 2) create a local maximum map of the contrast enhanced mammogram stored in the memory by performing a local maximum transform, wherein a 2-dimensional first filter implemented by the processor extracts a maximum pixel value from a region of interest (ROI); and 3) use the processor to derive a ratio map and store it in the memory, wherein the ratio map includes a pixel-wise ratio between the contrast enhanced mammogram and the local maximum map stored in the memory, wherein a pixel value of the ratio map is a measure of a relative response of the pixel of the ratio map to its local maximum in the raw mammogram.

Description

BRIEF DESCRIPTION OF THE FIGURES

[0032] FIG. 1 shows the normalising of a raw mammogram, where Normalised Image FIG. 1(d) is the pixel-wise ratio between the Contrast-Enhanced Mammogram FIG. 1(b) (i.e. contrast enhanced version of the Raw Mammogram shown as FIG. 1(a)) and the Local maximum Map FIG. 1(c), where FIG. 1(c) is derived by filtering FIG. 1(b) using a local maximum transform, which is implemented with a 2-dimensional first filter.

[0033] FIG. 2 shows two pairs of images: raw mammogram FIG. 2(a) and tomosynthesis FIG. 2(b) produced via imaging equipment from the same manufacturer (top row); raw mammograms from the same modalities but different manufacturers Siemens FIG. 2(c) and Philips FIG. 2(d); the ratio map normalisation (bottom row) removes the distinction between each raw image pairs, and further enhances the tissue structure.

[0034] FIG. 3 shows BACs extraction using Hessian-based multiscale filtering. A mammogram is normalised to a ratio map FIG. 3(a); a customised second filter is applied to generate a BACs probability image FIG. 3(b); binary BACs prediction FIG. 3(c) (in blue) is obtained by thresholding FIG. 3(b). The corresponding ground truth mask (in red) is shown in (d).

[0035] FIG. 4 shows a raw image (a) and its breast region segmentation (b). Here, the raw image is monochrome 2, so its pixel values in the breast region are smaller than the pixel values in the background region.

[0036] FIG. 5 shows the application of pixel saturation. Among the scaled pixels (see scaling Eq. (1)) in the breast region, the low and high saturation thresholds are determined by the top and bottom 1% of these pixels in (a). The saturation thresholds are applied to the entire image as in (b), yielding a saturated image (c).

[0037] FIG. 6 shows gamma correction where the correction maps low saturation to low out, and high saturation to high out following a nonlinear transform as shown. Gamma correction as applied to FIG. 5(c) yields FIG. 6(b).

[0038] FIG. 7 shows a local max transform. Taking a 3×3 image region (a) as an example, a local maximum transform implemented by the first filter extracts the maximum pixel value from this region. The first filter replaces the centre pixel value with the regional maximum as seen in (b). Local maximum transform on FIG. 6(b) yields FIG. 7(c).

[0039] FIG. 8 shows a ratio image derived from a pixel-wise ratio between FIG. 6(b) and FIG. 7(c) where the pixel values outside of the breast region are set to zero.

[0040] FIG. 9 shows workflow of normalizing a raw mammogram to a ratio image. The ratio image (f) is the pixel-wise ratio between (c) gamma corrected image and (d) local maximum image. The former is obtained from gamma correction of a 2% saturated image (b), and the latter is derived by filtering (c) using a local maximum first filter.

DETAILED DESCRIPTION OF THE INVENTION

[0041] In an embodiment, the method and system to differentiate and segment BACs from mammograms comprises three steps as given below.

1) Contrast enhancement for improved visibility of the breast tissue composition, whereby a region of the breast is segmented and a contrast-stretching algorithm applied to the segmented region. The algorithm maps the pixel values in the breast region to new values such that a portion of the data is saturated at low and high intensities of the input data.
2) Local ‘maximum’ transform, whereby a 2-dimensional first filter is designed to extract the maximum pixel value from a region of interest (ROI). An image of m rows and n columns can be broken into m×n overlapped ROIs, where each is centred at a pixel location (m, n). As a result, the first filter converts an image to another representation with the image size unchanged. Application of the first filter to the enhanced raw mammogram yields an image where its pixel represents a local maximum pixel value from the neighbouring pixels in the enhanced raw mammogram.
3) Ratio map derivation whereby the pixel value of the ratio map measures a relative response of the said pixel to its local maximum. Such relativity captures the difference between breast composition regardless of mammogram variations. As the pixel value of a ratio map ranges from 0 to 1, a rescaling to an 8-bit or 16-bit integer obtains a normalised image.

[0042] The pixel value in the normalised image represents a relative magnitude between this pixel and its neighbours in the raw image regardless of imaging conditions. Moreover, the normalisation amplifies the distinction between breast composition, so the BACs are in better contrast against other breast tissue. As illustrated in FIG. 1, the visibility of BACs is significantly improved in the normalised image (FIG. 1(d)) compared to the raw mammogram (FIG. 1(a)). Such clarity facilitates the development of a BAC detection algorithm.

[0043] Using Frangi's vessel enhancement second filter, tuned to accommodate the characteristics of BACs and a mammogram, the multiscale second-order local structure is examined (i.e. Hessian matrix) and a measure of vessel-probability is obtained from the eigenvalue of the Hessian matrix. The second filter enhances the tubular and elongated structures in the ratio map, which correspond to the BACs patterns.

[0044] With reference to FIG. 3, the image normalisation (FIG. 3(a)) amplifies the image gradient, in turn enhancing the contrast of the BACs with surrounding tissue. Frangi's vessel enhancement second filter is applied (FIG. 3(a)), yielding a BAC probability image (FIG. 3(b)). A self-adaptive thresholding algorithm then extracts the final BAC mask from the filtered image (FIG. 3(c)). The extracted BACs (FIG. 3(c)) are in good agreement with ground truth as manually marked by an experienced radiologist (FIG. 3(d)). Origin of ground truth is objective, provable data.

[0045] Direct application of Frangi's vessel enhancement second filter on the raw mammogram is not able to enhance the BACs structure in contrast to other breast tissues, which further fail the thresholding algorithm to segment BACs.

[0046] A further illustrative example of the three steps in the method is given below.

1) Contrast Enhancement

[0047] a) the breast region is segmented as follows: [0048] b) from raw mammogram image FIG. 4(a), a binary image is derived as seen in FIG. 4(b) where non-zero pixels indicate the location of the breast in the corresponding raw image. [0049] c) the pixel values inside the breast region in the raw image are scaled to a range between 0 and 1 followed by a logarithm transform using equation (1)

Scaled Raw Image=log(Raw Image/65535)  (1) [0050] d) image saturation thresholds are determined, i.e. guided by the segmentation FIG. 4(b), scaled pixels in the breast region of the raw image FIG. 4(a) are sorted in ascending order based on pixel values. Counting these pixels from higher to lower values, the top and bottom 1% of the pixels are determined as saturation thresholds. An illustration of determining saturation thresholds from FIGS. 4(a) and 4(b) is shown in FIG. 5(a). In the raw image, pixel values smaller than the low saturation threshold are set to low saturation threshold, and pixel values greater than the high saturation threshold are set to high saturation threshold. The saturated image has a pixel value range between low and high saturation thresholds as shown in FIG. 5(b). For example, FIG. 5(c) is the saturated image of FIG. 4(a).

[0051] e) Gamma correction maps the pixel values of a saturated image to a nonlinear range between specific bottom and top output thresholds following equation (2) and FIG. 6(a). The bottom and top output thresholds are set to parameters that yield the best contrast between fatty and dense (fibroglandular) tissue for example 0.01 and 1 respectively, and gamma correction factor is set to 0.5. After gamma correction, FIG. 5(c) becomes FIG. 6(b).

[00001]$\begin{array}{cc}\mathrm{Gamma}\mathrm{Corrected}\mathrm{Image}={\left[\frac{I-{\mathrm{low}}_{\mathrm{in}}}{hig{h}_{in}-{\mathrm{low}}_{\mathrm{in}}}\right]}^{\gamma }*\left(hig{h}_{out}-{\mathrm{low}}_{out}\right)+{\mathrm{low}}_{out}& \left(2\right)\end{array}$ [0052] where I represents the saturated image pixel values; low.sub.in and high.sub.in are low and high saturation thresholds respectively as seen in FIGS. 5(a) and 6(a), low.sub.out and high.sub.out are the specific bottom and top output thresholds as ‘low out’ and ‘high out’ in FIG. 6(a), and γ controls a weight between low out and high out.

2) Local Maximum Transform:

[0053] A 2-dimensional first filter is designed to extract the maximum pixel value from a region of interest (ROI) of 5 mm by 5 mm. Applying this first filter to the gamma corrected image yields an output where its pixel represents the maximum pixel value from the ROI centring on this pixel location. A demonstration of such local maximum transform is shown in FIGS. 7(a) and 7(b), and the transformed FIG. 6(b) is shown in FIG. 7(c).

3) Ratio Map Derivation:

[0054] A ratio map is used here to describe the pixel-wise ratio between the gamma corrected raw image and its local maximum map. For example, FIG. 8 is the result of FIG. 6(b) (pixel-wise) divided by FIG. 7(c). After scaling the floating-point ratio map to a 16-bit image, a final normalised image can be obtained. The entire workflow of deriving normalised ratio image from raw image is abstracted in FIG. 9.

[0055] The pixel value in the normalised image represents a relative magnitude between this pixel and its neighbors in the raw image, regardless of imaging conditions. As seen in FIG. 2, the proposed normalisation algorithm mitigates the difference between mammograms from different modalities and manufacturers. Furthermore, the normalisation naturally amplifies the image gradient, in turn enhancing the breast tissue with sharp edges. This may further facilitate the extraction of important tissue features, such as breast arterial calcification.

[0056] This invention has been described by way of example only, modifications and alternatives will be apparent to those skilled in the art. All such embodiments and modifications are intended to fall within the scope of the claims.