Systems and Methods For Predicting Lung Cancer Immune Therapy Responsiveness Using Quantitative Textural Analysis

Abstract

Methods and apparatus for predicting responsiveness to immune therapy in lung cancer. The method includes the steps of: identifying a first population of known responders and a second population of known non-responders; processing imaging data for the first and second populations using quantitative textural analysis (QTA); generating, for each member of both populations, quantitative metrics using the QTA; performing logistical regression on the quantitative metrics for both populations to yield a predictive signature expressed in the form of Y=Mx+B where x comprises mean pixel density; performing QTA on a lung cancer scan for a subsequent patient; comparing the predictive signature to one or more relevant metrics associated with the subsequent patient; and predicting responsiveness to immune therapy for the subsequent patient based on the comparison.

Claims

1. A biomarker signature for use in predicting responsiveness to immune therapy in lung cancer, expressed in the form
Y=Mx+B; where Y is a predictive indicator ranging from 0 to 1; B is a constant; M is a coefficient; and x indicates mean pixel density.

2. The signature of claim 1, wherein mean pixel density is derived from imaging data associated with a subjectively determined region of interest (ROI) surrounding a lung lesion.

3. The signature of claim 2, wherein mean pixel density is a measure of the average pixel density within a cluster of pixels derived from imaging data surrounding a tumor.

4. The signature of claim 3, wherein the imaging data comprises one of MRI, US, PET, DEXA, digital mammography, JPEGS, Angiography, SPECT, and gamma camera data.

5. The signature of claim 2, wherein the imaging data comprises CT scan data.

6. The signature of claim 1, wherein B has a value in the range of 0.1 to 1, and M has a value in the range of 0.001 to 0.1.

7. The signature of claim 5, wherein B has a value in the range of 0.4 to 0.6, and M has a value in the range of 0.007 to 0.09.

8. The signature of claim 5, wherein B has a value of about 0.5211, and M has a value of about 0.00858.

9. The signature of claim 6, derived using quantitative textural analysis (QTA) and logistical regression analysis on a first population of known responders and a second population of known non-responders.

10. A method of predicting responsiveness to immune therapy in lung cancer, comprising the steps of: identifying a first population of known responders and a second population of known non-responders; processing imaging data for the first and second populations using quantitative textural analysis (QTA); generating, for each member of both populations, quantitative metrics using the QTA; performing logistical regression on the quantitative metrics for both populations to yield a predictive signature expressed in the form of Y=Mx+B where x comprises mean pixel density; performing QTA on a lung cancer scan for a subsequent patient; comparing the predictive signature to one or more relevant metrics associated with the subsequent patient; and predicting responsiveness to immune therapy for the subsequent patient based on the comparison.

11. The method of claim 10, wherein: the imaging data comprises CT scan data; Y is a predictive indicator ranging from 0 to 1; B is a constant having a value in the range of 0.1 to 1; and M is a coefficient having a value in the range of 0.001 to 0.1.

12. The method of claim 11, wherein the step of processing using QTA comprises generating a histogram of the frequencies of occurrence (Y-axis) of pixels within discrete density boundaries (X-axis).

13. The method of claim 10, wherein the step of processing using QTA comprises using a spatial scale filters SSF3.

14. The method of claim 10, wherein the step of generating, for each member of both populations, quantitative metrics using the QTA comprises identifying a region of interest (ROI) surrounding a tumor.

15. The method of claim 10, wherein the first population of known responders is determined based on whether, in response to immune therapy, their corresponding imaging data reflects a simple reduction in at least 2 of: tumor size, tumor volume, and tumor density.

16. The method of claim 15, wherein the first population of known responders is further determined based on whether, in response to immune therapy, a reduction in tumor size coupled with negative trending growth kinetics in at least two of tumor size, tumor volume, and tumor density.

17. The method of claim 16, wherein the first population of known responders is further determined based on whether, in response to immune therapy, a reduction in tumor size coupled with negative trending growth kinetics in each of tumor size, tumor volume, and tumor density.

18. The method of claim 10, further comprising performing logistical regression on at least one of the following metrics: smoking history, gene mutation load, tumor markers, patient age, and patient gender.

19. The method of claim 18, wherein the quantitative metrics comprise mean pixel density, standard deviation of a histogram curve, mean positive pixel value of the pixels that are in the positive value range, entropy, skewness, and kurtosis.

20. Computer code stored in a non-transient medium which, when executed by a computer processor, performs the steps of: processing imaging data for a first population of responders and a second population of non-responders using quantitative textural analysis (QTA); generating, for each member of both populations, quantitative metrics using the QTA; and performing logistical regression on the quantitative metrics for both populations to yield a predictive signature expressed in the form of Y=Mx+B; where Y is a predictive indicator ranging from 0 to 1; B is a constant; M is a coefficient; and x indicates mean pixel density.

Description

BRIEF DESCRIPTION OF THE DRAWING FIGURES

[0016] Exemplary embodiments will hereinafter be described in conjunction with the appended drawing figures, wherein like numerals denote like elements, and:

[0017] FIG. 1 is an exemplary CT scan slice illustrating a region of interest (ROI) in accordance with various embodiments;

[0018] FIG. 2 is an exemplary histogram curve in accordance with various embodiments;

[0019] FIG. 3 is an exemplary data matrix corresponding to the histogram of FIG. 2 in accordance with various embodiments;

[0020] FIG. 4 is an exemplary histogram illustrating quantitative metrics in accordance with various embodiments;

[0021] FIG. 5 is an exemplary receiver operator characteristics (ROC) curve in accordance with various embodiments;

[0022] FIG. 6 is a first exemplary multiple regression output table in accordance with various embodiments; and

[0023] FIG. 7 is an alternate exemplary multiple regression output table in accordance with various embodiments.

DETAILED DESCRIPTION OF PREFERRED EXEMPLARY EMBODIMENTS

[0024] The following detailed description of the invention is merely exemplary in nature and is not intended to limit the invention or the application and uses of the invention. Furthermore, there is no intention to be bound by any theory presented in the preceding background or the following detailed description.

[0025] Various embodiments of the present invention relate to methods for developing a biomarker signature for predicting immune therapy responsiveness in lung cancers, including the steps of: i) obtaining cross sectional images from CT, MRI, US, PET, DEXA, Digital Mammography, JPEGS, Angiography, SPECT, gamma cameras, and/or optical platforms; ii) loading the imaging data into a suitable QTA platform (e.g., TexRAD) and selecting a region of interest (ROI) surrounding the tumor in the form of a rectangle, Ellipse, polygon, seed point, or other region encompassing the tumor; iii) selecting either a single slice or multiple slices for QTA; iv) selecting an appropriate filter algorithm (e.g., Liver, Lung, Mammo general, Mammo fine); v) filtering the pixels to a single common size and shape and clustering them together as nearest neighbors into groups of 2, 3, 4, 5, and 6 pixels each representing SSFs of 0 (no filter), 2 (fine), 3-4 (moderate), and 5-6 (coarse); vi) applying the different SSFs to the ROI area pixels and generating a histograph frequency curve for each SSF; vii) deconstructing each curve to yield metrics representing, for example, mean pixel density, standard deviation of the histogram curve, mean positive pixel value of the pixels that are in the positive value range, entropy, skewness, and kurtosis; viii) displaying the values in a matrix or otherwise representing the values in the form of equations; viii) performing logistical regression on the matrix values; and ix) using the results of the logistical regression individually or in combination with other clinical, laboratory, imaging, demographic, or other bio-informatic measurements to create imaging phenotypes for further connectivity to a predicted outcome.

[0026] In a preferred embodiment, a volumetric CT imaging data set for each of a plurality of lung cancer tumors is analyzed. Referring now to FIG. 1, for each slice 100 in each data set, a region of interest (ROI) 102 is identified (typically manually). For each data set, an optimum slice is selected for further processing, such as the most heterogeneous, irregular slice from the data set. Alternatively, the entire data set or a subset thereof may be employed.

[0027] For the selected slice, the pixels within the ROI 102 are processed using QTA. An initial processing step involves selecting, an appropriate filter based on thresholds of density, with air being the least dense and bone being the most dense. That is, the filter seeks to remove air and bone pixels, leaving only pixels within the ROI of biological relevance.

[0028] Referring now to FIG. 2, the ATQ platform then generates a histogram 200 of the frequencies of occurrence (Y-axis) of pixels within discrete density boundaries (X-axis). CT images typically display greyscale values in terms of Hounsfield units, where water=0 in a scale from 1500 to +1500. Pixels more dense than water are positive; pixels less dense than water have negative values. Density values may be generally grouped into four tissue types that exhibit contrast: air (less than 80); fat (80 to 20); water/soft tissue (20 to 300); and bone (above 300). After applying a selected band pass filter to remove the very high and very low density pixels, a histogram is generated upon which the following calculations and statistical analyses are performed.

[0029] Referring now to FIG. 3, the system (e.g., TexRAD) then calculates various metrics for one or more spatial scale filter (SSF) filter values 301 associated with the histogram, including: i) the mean pixel value 302 representing the average density within a cluster of pixels at a given SSF level; ii) the standard deviation (SD) 304 which is a measure of tumor heterogeneity and microstructural change; iii) entropy 306 representing the mean density of clustered pixels over the entire ROI area (Ln [mean density/total pixels]) and is based on different filtering parameters that are reflective of tumor homo/heterogeneity; iv) mean positive pixel (mpp) value 308, sometimes regarded as a measure of hypoxia; v) skewness 310 used to measure the symmetry of contrast distribution in regions of interest, where skewness is measured by the slant of the peak either to the right (negative skewness) or to the left (positive skewness); vi) kurtosis 312 which is determined by the height of the histogram and regarded as a measure of tumor angiogenesis, vascular shunting and/or tumor homogeneity; and vii) the total pixel number 314 for the histogram.

[0030] The foregoing histogram and the associated metrics embody biological information, which the present inventor seeks to harness in the form of a signature useful in predicting responsiveness to immune therapy for lung cancer. Specifically, the present inventor seeks to characterize the data in terms of a signature against which future patient scans may be evaluated to predict responsiveness to immune therapy with a high degree of confidence. FIG. 4 illustrates a histogram 400 and graphically depicts the following exemplary metrics: standard deviation 402; mpp 404; skewness 406; and kurtosis 408.

[0031] In this regard, immune therapy broadly involves using drugs to provoke the immune system to attack the cancer, rather than using drugs to attack the cancer directly. Immune therapy works well in approximately 30% of patients, but is quite expensive (e.g., $25,000/month). In approximately 70% of patients immune therapy helps only a little or not at all. It is therefore desirable to predict in advance whether a particular patient is likely to respond to immune therapy, based on comparison to a signature previously derived from a population known to respond to immune therapy.

[0032] After evaluating each SSF filter level (corresponding to 0, 2, 3, 4, 5, or 6 adjacent clustered pixels) independently, the associated metric values may be summarized in a matrix as shown in FIG. 3, which represents a deconstructed histogram for the pixels within an ROI of a single slice for a single patient for various SSF values.

[0033] The foregoing metrics may then be processed using a simple T-test to determine whether a difference between respective mean values for two population groups (e.g., responder and non-responder) is unlikely to have occurred because of random chance in sample selection. Each metric having a significant difference between the average value for the responder population and the average value for the non-responder population is a good candidate for including in the signature.

[0034] A more robust signature may be derived using logistical regression to yield a signature representative of the underlying biology, where metrics which influence the outcome (immune therapy responder or non-responder) are preserved in the model, and where metrics which do not influence the outcome are not preserved in the model. More particularly, known responders are allocated a 1 and known non-responders are allocated a zero, where zero and 1 are the dependent variables in the logistical regression analysis. The logistical regression model then reveals the principal factors that align with responders/non-responders, as well as their coefficients. This can be done using forward, backward, step wise, or any other desired statistical protocol.

[0035] In an embodiment, the logistical regression analysis employs a matrix of equations of the form 1=Ax.sub.1+Bx.sub.2+Cx.sub.3+Dx.sub.4+Ex.sub.5+Fx.sub.6 for responders, and of the form 0=Ax.sub.1+Bx.sub.2+Cx.sub.3+Dx.sub.4+Ex.sub.5+Fx.sub.6 for non-responders, where x.sub.1 corresponds to the mean, x.sub.2 corresponds to the standard deviation, x.sub.3 corresponds to entropy, x.sub.4 corresponds to mpp, x.sub.5 corresponds to skewness, and x.sub.6 corresponds to kurtosis. The logistical regression analysis then determines which metrics influence immune therapy responsiveness, and calculates the associated coefficients (e.g., A-F) for the metrics retained in the model.

[0036] In an alternate embodiment, one or more extra columns may be used in addition to the aforementioned metrics to enhance the predictive value of the signature. This additional column or columns may relate to bio-informatic metrics such as, for example, smoking history, gene mutation load, tumor markers, pathology information (age, gender); logistical regression analysis may then be performed on all columns.

[0037] The logistical regression process, which may be implemented in an algorithm, produces a master equation using well known techniques, and retains the statistically important independent variables and discards the statistically unimportant independent variables. A binary logistic model may be used to estimate the probability of a binary response based on one or more predictor (or independent) variables (features).

[0038] Once a preliminary logistical regression analysis has been performed and one or more metrics are identified as significant, they may be expressed as an equation of the form 1=Ax+By+Cz, where x, y, and z are the metrics determined to be linked to the responder outcome, and A, B, and C are their corresponding coefficients (at least one of which is non-zero). With momentary reference to FIG. 5, this equation may then be expressed as a receiver operator characteristics (ROC) curve 500 and analyzed to determine the cut-off value which yields the highest sensitivity. For example, by inspecting an exemplary receiver operator characteristics curve, one may conclude that if x, y, and/or z exceed predetermined threshold values, the patient is statistically likely to be a responder.

[0039] Once a linkage is established between responders and the metrics retained in the logistical regression model, the linkage is preferably validated before declaring the signature statistically stable. That is, if a new candidate is determined to be a responder based on comparison of that patient's scan metrics to the signature, we then follow up to confirm that he does in fact respond. If he does not respond when predicted to do so, the reasons underlying the discrepancy are addressed or the signature revised. Once the signature is positively validated, the validated signature becomes the predictive test we know to a statistical certainty that declared responders will in fact respond.

[0040] The results of a first logistical regression analysis using a sample size of 14 for predicting immune therapy responsiveness in lung cancers are tabulated in FIG. 6; an alternate logistical regression analysis using a sample size of 32 for predicting immune therapy responsiveness in lung cancers are tabulated in FIG. 7, which revealed filter level SSF3 was more statistically significant than other SSF levels for predicting immune response for lung cancers. In this regard, the data was not restricted to tumors located in the lungs but, rather, included all lung cancers regardless of the physical location of the tumor within or outside the lung.

[0041] In both sample sizes (14 and 32) the signatures retained only the mean; the remaining metrics (sd, entropy, mpp, skewness, and kurtosis) were not retained in the model. In the analysis shown in FIG. 7, the signature may be expressed as Y=0.5211+0.008576(x), where x is the independent variable and corresponds to the mean value from the QTA. The predictive value of this equation is self-evident; for all mean values greater than (2.46), there is a greater than 50% probability that the patient is a responder; for all mean values greater than (50.01), there is a greater than 95% probability that the patient is a responder, and so on.

[0042] In the analysis shown in FIG. 6, the signature may be expressed as Y=0.4546+0.007114(x), where x corresponds to the mean value from the QTA. The predictive value of this equation is also self-evident, and generally parallels the equation corresponding to FIG. 7.

[0043] Both equations suggest that immune responsiveness is most sensitive to the average (or mean) density of pixels within the region of interest for lung cancers.

[0044] In various embodiments, it may be desirable to bias the logistical regression analysis in the direction of responders, for example, by using a greater number of responder scans than non-responder scans in the data set. That is, since one objective is to define a responsive signature, it is appropriate to bias the data set to tend towards responders, or else the linkage between responders and the metrics influencing responders may be suppressed or washed out entirely. In various embodiments, the data set may comprise a ration of responders to non-responders in the range of 1:1 to 2:1.

[0045] The manner in which the responder population and the non-responder population are defined will now be described in accordance with various embodiments. After evaluating the pixels within an ROI for a scan slice (or group of slices) associated with a particular patient, immune therapy is administered to that patient. Another full volumetric scan is subsequently taken later in time (typically 4 or 8 weeks following introduction of the immune therapy drugs). Based on that subsequent volumetric data, it is determine whether the patient is a responder or a non-responder within the first 8 weeks (this defines an early responder). It remains to determine how to define whether a patient responds to immune therapy.

[0046] More particularly, shrinkage in tumor size is an important factor, but not sufficient alone because not all tumors shrink in responders; it is therefore appropriate to also consider changes in the volume and/or density of the tumor. Thus, in one embodiment a responder may be defined as a patient who, in response to immune therapy, experiences a simple reduction in at least 2 of: tumor size, volume, and density.

[0047] Moreover, it is known that sometimes tumors may not simply shrink; rather, tumors may exhibit tumor growth kinetics which characterize or define the rate at which the tumor size, volume, and/or density changes over time in the presence of immune therapy. Using a growth kinetics analysis, the mere fact a tumor has shrunk is not enoughit itselfto declare a patient to be a responder. Accordingly, in an alternative embodiment, a responder is conservatively defined where the tumor growth kinetics (in terms of size, volume, and/or density) also trend negative over time. That is, a simple reduction in tumor size is not sufficient to declare the patient a responder; rather, a reduction in tumor size coupled with negative trending growth kinetics (e.g., one, two, or three of a negative growth kinetic for tumor size, volume, and density) is required to conservatively define a responder. Using such a conservative definition for a responder greatly enhances the level of confidence in the predictive value of the signature, because responders which satisfy only the static reduction threshold but who do not also satisfy the kinetic reduction metric are not considered responders.

[0048] While the present invention has been described in the context of the foregoing embodiments, it will be appreciated that the invention is not so limited. For example, the various geometric features and chemistries may be adjusted to accommodate additional applications based on the teachings of the present invention.

[0049] A biomarker signature is thus provided for use in predicting responsiveness to immune therapy in lung cancer, expressed in the form Y=Mx+B; where Y is a predictive indicator ranging from 0 to 1; B is a constant; M is a coefficient; and x indicates mean pixel density.

[0050] In an embodiment, mean pixel density is derived from imaging data associated with a subjectively determined region of interest (ROI) surrounding a lung lesion.

[0051] In an embodiment, mean pixel density is a measure of the average pixel density within a cluster of pixels derived from imaging data surrounding a tumor.

[0052] In an embodiment, the imaging data comprises one of MRI, US, PET, DEXA, digital mammography, JPEGS, Angiography, SPECT, and gamma camera data.

[0053] In an embodiment, the imaging data comprises CT scan data.

[0054] In an embodiment, B has a value in the range of 0.1 to 1, and M has a value in the range of 0.001 to 0.1.

[0055] In an embodiment, B has a value in the range of 0.4 to 0.6, and M has a value in the range of 0.007 to 0.09.

[0056] In an embodiment, B has a value of about 0.5211, and M has a value of about 0.00858.

[0057] In an embodiment, the signature is derived using quantitative textural analysis (QTA) and logistical regression analysis on a first population of known responders and a second population of known non-responders.

[0058] A method is also provided for predicting responsiveness to immune therapy in lung cancer patients. The method includes: identifying a first population of known responders and a second population of known non-responders; processing imaging data for the first and second populations using quantitative textural analysis (QTA); generating, for each member of both populations, quantitative metrics using the QTA; performing logistical regression on the quantitative metrics for both populations to yield a predictive signature expressed in the form of Y=Mx+B where x comprises mean pixel density; performing QTA on a lung cancer scan for a subsequent patient; comparing the predictive signature to one or more relevant metrics associated with the subsequent patient; and predicting responsiveness to immune therapy for the subsequent patient based on the comparison.

[0059] In an embodiment, the imaging data comprises CT scan data; Y is a predictive indicator ranging from 0 to 1; B is a constant having a value in the range of 0.1 to 1; and M is a coefficient having a value in the range of 0.001 to 0.1.

[0060] In an embodiment, the step of processing using QTA comprises generating a histogram of the frequencies of occurrence (Y-axis) of pixels within discrete density boundaries (X-axis).

[0061] In an embodiment, the step of processing using QTA comprises using a spatial scale filters SSF3.

[0062] In an embodiment, the step of generating, for each member of both populations, quantitative metrics using the QTA comprises identifying a region of interest (ROI) surrounding a tumor.

[0063] In an embodiment, the first population of known responders is determined based on whether, in response to immune therapy, their corresponding imaging data reflects a simple reduction in at least 2 of: tumor size, tumor volume, and tumor density.

[0064] In an embodiment, the first population of known responders is further determined based on whether, in response to immune therapy, a reduction in tumor size coupled with negative trending growth kinetics in at least two of tumor size, tumor volume, and tumor density.

[0065] In an embodiment, the first population of known responders is further determined based on whether, in response to immune therapy, a reduction in tumor size coupled with negative trending growth kinetics in each of tumor size, tumor volume, and tumor density.

[0066] In an embodiment, the method further includes performing logistical regression on at least one of the following metrics: smoking history, gene mutation load, tumor markers, patient age, and patient gender.

[0067] In an embodiment, the quantitative metrics comprise mean pixel density, standard deviation of a histogram curve, mean positive pixel value of the pixels that are in the positive value range, entropy, skewness, and kurtosis.

[0068] Computer code is also provided. The computer code is stored in a non-transient medium which and, when executed by a computer processor, performs the steps of: processing imaging data for a first population of responders and a second population of non-responders using quantitative textural analysis (QTA); generating, for each member of both populations, quantitative metrics using the QTA; and performing logistical regression on the quantitative metrics for both populations to yield a predictive signature expressed in the form of Y=Mx+B; where Y is a predictive indicator ranging from 0 to 1; B is a constant; M is a coefficient; and x indicates mean pixel density.

[0069] As used herein, the word exemplary means serving as an example, instance, or illustration. Any implementation described herein as exemplary is not necessarily to be construed as preferred or advantageous over other implementations, nor is it intended to be construed as a model that must be literally duplicated.

[0070] While the foregoing detailed description will provide those skilled in the art with a convenient road map for implementing various embodiments of the invention, it should be appreciated that the particular embodiments described above are only examples, and are not intended to limit the scope, applicability, or configuration of the invention in any way. To the contrary, various changes may be made in the function and arrangement of elements described without departing from the scope of the invention.