Generalizable and interpretable deep learning framework for predicting MSI from histopathology slide images
A generalizable and interpretable deep learning model for predicting microsatellite instability from histopathology slide images is provided. Microsatellite instability (MSI) is an important genomic phenotype that can direct clinical treatment decisions, especially in the context of cancer immunotherapies. A deep learning framework is provided to predict MSI from histopathology images, to improve the generalizability of the predictive model using adversarial training to new domains, such as on new data sources or tumor types, and to provide techniques to visually interpret the topological and morphological features that influence the MSI predictions.
1. A computing device configured to generate an image-based microsatellite instability (MSI) prediction model, the computing device comprising one or more processors configured to:
obtain, via one or more processors, a set of stained histopathology images from one or more image sources, the set of stained histopathology images having a first cancer type-specific bias or a first data source-specific bias;
store in an electronic database, using the one or more computing devices, an association between the histopathology slide images and the plurality of MSI classification labels;
apply, via one or more processors, a statistical model to analyze the set of stained histopathology images and generate an initial baseline MSI prediction model, the initial baseline MSI prediction model exhibiting cancer type-specific bias or data source-specific bias;
apply, via one or more processors, an adversarial training to the baseline MSI prediction model; and
generate, via one or more processors, an adversarial trained MSI prediction model configured to predict MSI status for subsequent stained histopathology images,
the adversarial trained MSI prediction model characterized by a reduction in cancer type-specific bias or data source-specific bias in comparison to the initial baseline MSI prediction status model, and
the adversarial trained MSI prediction model configured to predict an MSI status including at least two of: (i) Microsatellite-High (MSI-H); (ii) Microsatellite-Stable (MSS); or Microsatellite-Low (MSI-L).
2. The computing device of claim 1 , wherein the statistical model is a Neural Network or Support Vector Machine (SVM).
3. The computing device of claim 1 , wherein the one or more processors are configured to:
obtain at least one of the subsequent stained histopathology images;
apply the adversarial trained MSI prediction model to the at least one subsequent stained histopathology image and predict MSI status;
examine the at least one subsequent stained histopathology image and identify patches associated with the MSI status; and
generate a guided backpropagation histopathology image from the at least one subsequent stained histopathology image, the guided backpropagation histopathology image depicting the patches associated with the MSI status.
4. The computing device of claim 3 , wherein the patches comprise pixels or groups of pixels.
5. The computing device of claim 3 , wherein the patches are characterized by topology and/or morphology of pixels or groups of pixels.
6. The computing device of claim 3 , wherein the one or more processors are configured to examine the at least one subsequent stained histopathology image and identify the patches associated with the MSI status using a gradient-weighted class activation map.
7. The computing device of claim 1 , wherein the statistical model is a machine learning process.
8. A computer-implemented method to generate an image-based microsatellite instability (MSI) prediction model, the method comprising:
obtaining, via one or more processors, a set of stained histopathology images from one or more image sources, the set of stained histopathology images having a first cancer type-specific bias or a first data source-specific bias;
storing in an electronic database, using the one or more computing devices, an association between the histopathology slide images and the plurality of MSI classification labels;
applying, via one or more processors, a statistical model to analyze the set of stained histopathology images and generating an initial baseline MSI prediction model, the initial baseline MSI prediction model exhibiting cancer type-specific bias or data source-specific bias;
applying, via one or more processors, an adversarial training to the initial baseline MSI prediction model; and
generating, via one or more processors, an adversarial trained MSI prediction model configured to predict MSI status for subsequent stained histopathology images,
the adversarial trained MSI prediction model characterized by a reduction in cancer type-specific bias or data source-specific bias in comparison to the initial baseline MSI prediction status model, and
the adversarial trained MSI prediction model configured to predict an MSI status including at least two of: (i) Microsatellite-High (MSI-H); (ii) Microsatellite-Stable (MSS); or Microsatellite-Low (MSI-L).
9. A computing device configured to generate an image-based microsatellite instability (MSI) prediction model, the computing device comprising one or more processors configured to:
obtain, via one or more processors, a set of stained histopathology images from one or more image sources, the set of stained histopathology images having a first cancer type-specific bias or a first data source-specific bias;
store, via one or more processors, in an electronic database, using the one or more computing devices, an association between the histopathology slide images and the plurality of MSI classification labels; and
apply, via one or more processors, a statistical model to analyze the set of stained histopathology images; and
generate, via one or more processors, an adversarial trained MSI prediction model configured to predict MSI status for subsequent stained histopathology images,
the adversarial trained MSI prediction model characterized by a reduction in cancer type-specific bias or data source-specific bias in comparison to the initial baseline MSI prediction status model, and
the adversarial trained MSI prediction model configured to predict an MSI status including at least two of: (i) Microsatellite-High (MSI-H); (ii) Microsatellite-Stable (MSS); or Microsatellite-Low (MSI-L).
10. The computing device of claim 9 , wherein the statistical model is a Neural Network or Support Vector Machine (SVM).
11. The computing device of claim 9 , wherein the one or more processors are configured to:
obtain at least one of the subsequent stained histopathology images;
apply the trained MSI prediction model to the at least one subsequent stained histopathology image and predict MSI status;
examine the at least one subsequent stained histopathology image and identify patches associated with the MSI status; and
generate a guided backpropagation histopathology image from the at least one subsequent stained histopathology image, the guided backpropagation histopathology image depicting the patches associated with the MSI status.
12. The computing device of claim 11 , wherein the patches comprise pixel or groups of pixels.
13. The computing device of claim 11 , wherein the patches are characterized by topology and/or morphology of pixels or groups of pixels.
14. The computing device of claim 11 , wherein the one or more processors are configured to examine the at least one subsequent stained histopathology image and identify the patches associated with the MSI status using a gradient-weighted class activation map.
15. The computing device of claim 9 , wherein the statistical model is a machine learning process.