Systems and methods for assessing liver pathology
In some aspects, the described systems and methods provide for a method for training a deep learning model to assess liver pathology, including accessing annotated liver pathology images associated with a group of patients in one or more randomized controlled clinical trials of nonalcoholic steatohepatitis therapy, each of the annotated liver pathology images including at least one annotation describing one or more tissue characteristic categories for a portion of the image, and training the deep learning model based on the annotated liver pathology images to predict the tissue characteristic categories, selected from a group comprising steatosis, lobular inflammation, hepatocyte ballooning, and fibrosis stage.
1. A system for training a deep learning model to assess liver pathology, the system comprising:
at least one computer hardware processor; and
at least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by the at least one computer hardware processor, cause the at least one computer hardware processor to perform:
accessing a plurality of annotated liver pathology image associated with a group of patients in one or more randomized controlled clinical trials of nonalcoholic steatohepatitis therapy, wherein each of the plurality of annotated liver pathology images includes at least one annotation describing one or more tissue characteristic categories for a portion of the image;
training the deep learning model based on the plurality of annotated liver pathology images to predict the one or more tissue characteristic categories for a liver pathology image, wherein the one or more tissue characteristic categories are selected from a group comprising steatosis, lobular inflammation, hepatocyte ballooning, and fibrosis stage;
storing the trained deep learning model on at least one storage device; and
for a liver pathology image, evaluating a performance of the deep learning model by comparing a fraction of tissue area in the liver pathology image assigned to each NAS component with an ordinal score determined by a pathologist, wherein the ordinal score is determined based on the NAS components, wherein intra-observer reproducibility of the ordinal score of the liver pathology image by multiple pathologists is lower than reproducibility of predictions from multiple iterations of the trained deep learning model processing the liver pathology image,
wherein a group comprising steatosis, lobular inflammation, and hepatocyte ballooning corresponds to components of a nonalcoholic fatty liver disease activity score (NAS).
2. The system of claim 1 , wherein the processor-executable instructions cause the at least one computer hardware processor to further perform:
measuring a feature at or prior to the therapy of a patient, and measuring the feature at a time point following therapy of the patient and calculating differences in feature values to determine treatment-dependent effects.
3. The system of claim 1 , wherein the processor-executable instructions cause the at least one computer hardware processor to further perform:
processing, using the trained deep learning model, a liver pathology image to generate predicted one or more tissue characteristic categories for the liver pathology image; and
generating, for the liver pathology image, based on the predicted one or more tissue characteristic categories, a slide-level continuous score that corresponds to an underlying heterogeneous pattern of fibrosis observed in the liver pathology image.
4. The system of claim 3 , wherein the one or more tissue characteristic categories are predicted on a pixel-level basis.
5. The system of claim 3 , wherein the slide-level continuous score represents a continuum of severity of cellular and stromal injury to capture disease heterogeneity among different patients, wherein an ordinal classification system does not capture the disease heterogeneity among the different patients.
6. The system of claim 1 , wherein the processor-executable instructions cause the at least one computer hardware processor to further perform:
extracting values for a plurality of complex features from each of the annotated liver pathology images, wherein the plurality of complex features includes one or more features not included in the NAS components; and
training the deep learning model based on the plurality of annotated liver pathology images and the values for the plurality of complex features.
7. The system of claim 6 , wherein the plurality of complex features comprises at least one ratio of the NAS components, including a ratio of area of hepatocellular ballooning to steatosis.
8. The system of claim 6 , wherein the one or more features not included in the NAS components include portal inflammation and bile duct.
9. The system of claim 1 , wherein the processor-executable instructions cause the at least one computer hardware processor to further perform:
determining a score for a response metric that captures change in fibrosis pattern before and following the therapy by computing a distance between distribution of fibrosis stages at baseline and following the therapy.
10. The system of claim 9 , wherein the processor-executable instructions cause the at least one computer hardware processor to further perform:
based on a threshold, classifying patients as responder or non-responder to the therapy based on the score for the response metric for a respective patient.
11. The system of claim 1 , wherein the deep learning model comprises one or more deep convolutional neural networks.