Deep Learning Identifies Microsatellite Instability in H&E Whole Slide Images from Prostate, Esophageal, and Gastric Cancers and Generalizes across Cancer Types

USCAP 2022 Annual Meeting, Tempus-authored
Authors Rohan Joshi, Andrew Kruger, Elle Moore, Ryan Jones, Martin Stumpe

Background: Defective mismatch repair (dMMR) proteins and high microsatellite instability (MSI-H) are associated with a positive response to checkpoint inhibitor therapy in colorectal and non-colorectal cancers. Because of low prevalence in certain noncolorectal cancers, testing for dMMR and/or MSI-H is not routinely performed, particularly at the time of initial biopsy. Assaying would be more practical if a method existed to enrich for likely-positive patients. Here, we tested the ability to predict MSI status from H&E whole slide images (WSIs) in prostate cancer and to generalize MSI prediction to gastric and esophageal cancers.

Design: WSIs and MSI-H—or microsatellite stable (MSS)—labels (obtained by next-generation sequencing) were collected from primary or metastatic colorectal, endometrial, and prostate cancer specimens (N=3296, Figure 1). An attention-based convolutional neural network was trained to predict the MSI label from each WSI in the training set. Hyperparameters and operating points were selected using the optimization set of data, targeting high sensitivity for prostate cancer prediction (prostate cancer model) or simultaneous prostate, endometrial, and colorectal cancer prediction (gastric/esophageal cancer model). Results are reported on a fully-independent holdout set (prostate cancer model) or independently collected datasets (gastric and esophageal cancer model).

Results: We trained two different models and assessed accuracy across three datasets, finding comparable sensitivity/specificity between models trained and tested on the same cancer type vs different types (Table 1). Using estimated real-world prevalences of 5%, 2%, and 20% of MSI-H (in prostate, esophageal, and gastric cancers, respectively) and a positive MSI-H prediction from the models, we expect that 15%, 6%, and 32% (respectively) of patients would have detectable MSI-H status on follow-up testing after the model result.

Conclusions: H&E WSIs contain features that can predict the MSI-H phenotype in prostate, esophageal, and gastric cancers. A deep learning model showed discriminative ability on esophageal and gastric cancers despite being trained on other cancers, suggesting generalization across cancer types. Enrichment of patients likely to be MSI-H using this approach could make MSI or dMMR testing more feasible for routine use in these cancers. Future work includes expansion of the study to the institution’s entire set of data and examination of potential model biases.


VIEW THE PUBLICATION – Abstract no. 992