Methods of assessing diseases using image classifiers
The present disclosure provides methods and systems using machine learning to assess one or more of a patient's biomarkers to analyze various conditions, including cancers, such as breast cancer. The present systems and methods can be trained to analyze patient's biomarker data to form prognoses, diagnoses, and treatment suggestions. Further, the present systems and methods can use biomarker feature data and clinical feature data to create novel correlations in order to provide more accurate, patient-specific diagnoses, prognoses, and treatment suggestions.
1. A method for assessing cancer risk, the method comprising:
training a first image classifier to correlate biomarker data from tissue samples to features in images of the tissue samples using a first set of training data;
providing an image of a tissue sample from a subject to the first image classifier;
analyzing the image using the first image classifier to generate a first cancer risk score;
training at least a second image classifier to correlate biomarker data from tissue samples to features in images of the tissue samples using at least a second set of training data;
wherein if the first cancer risk score does not surpass a threshold, analyzing the image using the at least second image classifier to generate at least a second cancer risk score; and
classifying a level of cancer risk for the subject using the first and at least second cancer risk scores.
2. The method of claim 1 , wherein:
the first set of training data comprises biomarker data and tissue sample images correlated with a cancer risk between a high and low cancer risk; and
the second set of training data consists of biomarker data and tissue sample images correlated with at least a high and/or low cancer risk.
3. The method of claim 2 , wherein the first set of training data further comprises biomarker data and tissue sample images correlated with a high and/or low cancer risk.
4. The method of claim 2 , wherein the threshold is surpassed when the first cancer risk score indicates a high or low cancer risk.
5. The method of claim 4 , wherein the method further comprises:
training one or more additional image classifiers to correlate biomarker data from tissue samples to features in images of the tissue samples using at least a different set of training data for each classifier; and
wherein if the first cancer risk score does not surpass a threshold, analyzing the image using the additional image classifiers to generate additional cancer risk scores.
6. The method of claim 5 , wherein classifying includes using the first, second, and additional cancer risk scores.
7. The method of claim 4 , wherein the cancer risk scores are assigned to a group, wherein each group corresponds to a different range of cancer risk scores.
8. The method of claim 7 , wherein the second set of training data consists of biomarker data and tissue sample images correlated with a cancer risk score within the cancer risk score range of one or more selected group.
9. The method of claim 8 , wherein the selected groups consist of a group with a highest range of cancer risk scores and a group with a lowest range of cancer risk scores.
10. The method of claim 7 , wherein the method further comprises:
training one or more additional image classifiers to correlate biomarker data from tissue samples to features in images of the tissue samples using at least a different set of training data for each classifier; and
wherein if the first cancer risk score does not surpass a threshold, analyzing the image using the additional image classifiers to provide additional cancer risk scores.
11. The method of claim 10 , wherein classifying includes using the first, second, and additional cancer risk scores.
12. The method of claim 11 , wherein each of the second and additional classifiers are each trained using a different data set, and each different data set consists of biomarker data and tissue sample images correlated with a cancer risk score falling within the cancer risk score range of a different group.
13. The method of claim 11 , wherein there are five groups, and the groups consist of:
a first group with a range cancer risk scores correlated with an ultra-high risk of cancer;
a second group with a range of cancer risk scores correlated with a high risk of cancer;
a third group with a range of cancer risk scores correlated with an ambiguous risk of cancer;
a fourth group with a range of cancer risk scores correlated with a low risk of cancer; and
a fifth group with a range of cancer risk scores correlated with an ultra-low risk of cancer.
14. The method of claim 13 , wherein if the first cancer risk score falls within the range of the third group, the first cancer score does not surpass the threshold.
15. The method of claim 14 , wherein each of the second and additional classifiers are each trained using a different data set, and each different data set consists of biomarker data and tissue sample images correlated with a cancer risk score falling within the first, second, third, fourth, or fifth group.
16. The method of claim 15 , wherein training the second and additional classifiers further includes:
dividing each of the groups into at least two subdivisions, each subdivision corresponding to a different range of cancer risk scores; and
training each of the second and additional classifiers using biomarker data and tissue sample images correlated with a cancer risk score within one or more of the subdivisions.
17. The method of claim 1 , further comprising:
segmenting the tissue image from the subject into segments;
providing the segments to the image classifiers for analysis to generate cancer risk scores; and
classifying a level of cancer risk for the subject using the generated cancer risk scores for each segment.
18. The method of claim 1 , wherein the biomarker data comprises genomic data and/or RNA expression data.
19. The method of claim 18 , wherein RNA expression data comprises RNA expression levels determined using sequencing or microarray analysis.
20. The method of claim 19 , wherein RNA expression levels are determined by interrogating a sample with probes specific for a panel of genes and measuring expression levels for positive probe responses.
21. A method for assessing cancer risk, the method comprising:
providing an image of a tissue sample from a patient to an image classifier trained on a plurality of tissue images, each associated with a cancer risk score;
operating the image classifier on the image to output a risk score for the patient;
when the risk score is within a range of threshold values between high-risk and low-risk scores, providing the image or the tissue sample to a second classifier that has been trained on tissue images, genomic characteristics, phenotypic characteristics or biomarker characteristics to obtain a second risk score for the patient.
22. The method of claim 21 , wherein the cancer risk scores associated with the plurality of tissue images are divided into groups that include at least a highest risk group, a middle group for the range of threshold values between high-risk and low-risk scores, and a lowest risk group.
23. The method of claim 22 , wherein the selected subset of the plurality of tissue images does not include any images associated with scores in the middle group.
24. A method for assessing cancer risk, the method comprising:
providing an image of a tissue sample from a patient to an image classifier trained on a plurality of tissue images, each associated with a cancer risk score;
operating the image classifier on the image to output a risk score for the patient;
when the risk score is within a range of threshold values between high-risk and low-risk scores, providing the image to a second classifier that has been trained using a selected subset of the plurality of tissue images to obtain a second risk score for the patient.
25. The method of claim 24 , wherein the image classifier is trained by CLustering-constrained Attention Multiple instance (CLAM) learning.
26. The method of claim 24 , wherein the image is a digital whole-slide image (WSI) of a stained section of tissue on a slide.
27. The method of claim 26 , wherein the section includes tumor tissue and healthy stroma from the patient.
28. The method of claim 24 , wherein the cancer risk score associated with each of the plurality of tissue images was obtained by genomic analysis for expression of a panel of genes.
29. The method of claim 28 , wherein the genomic analysis is selected from arrays, next generation sequence, PCR, RNAseq, and hybrid capture.
30. The method of claim 24 , wherein the second classifier has been trained by including only those images of the plurality of tissue images associated with a very high or very low risk score.
31. The method of claim 24 , wherein the image classifier operates using a neural network or random forest.
32. The method of claim 24 , wherein the image classifier divides the image into segments, scores each segment, identifies highest- and lowest-scoring segments, and calibrates scores for remaining segments based on the highest- and lowest-scoring segments.
33. The method of claim 24 , wherein the cancer risk scores associated with the plurality of tissue images are divided into a number of groups such that a cutoff value between a positive and a negative cancer risk score lies within one group and not an a boundary between two groups.
34. The method of claim 33 , wherein the selected subset of the plurality of tissue images does not include any images associated with scores in the one group.
35. The method of claim 24 , further comprising pre-processing the image before operating the image classifier to select and include regions containing only certain tissue types.
36. The method of claim 35 , wherein the image classifier computes an attention map for the included regions.
37. The method of claim 24 , wherein the image classifier is trained on a plurality of tissue images that includes at least ten thousand images.