Abstract 5012: Deep Gaussian process with uncertainty estimation Improves microsatellite instability prediction based on whole slide image: A retrospective multicenter and multiethnic cohort study

Abstract

Abstract Background: Microsatellite Instability (MSI) is a prognostic and predictive biomarker which can guide treatments including immunotherapy for patients with gastrointestinal cancer. Universal MSI testing would benefit to patients to receive a better therapy, but many patients remain untested. To deliver broadly accessible MSI testing, deep learning models have demonstrated their feasibility for MSI prediction using H&E-stained whole-slide images (WSIs). However, to be deployed into clinical routine care, prediction models should be validated in datasets from multicenter and multiethnic groups. In addition, these models should provide uncertainty of prediction to help clinicians to make informed decisions. Method: We develop a MSI prediction model using WSIs based on Deep Gaussian process (DGP), a Bayesian model being able to model the uncertainty in prediction. We implement a DGP model in the transfer learning, where a WPI is decomposed into multiple non-overlap image patches which are converted to feature vectors by a pretrained convolutional neural network (CNN). Then, the DGP model makes a prediction for an image by averaging score function values at all the feature vectors in the image in weakly supervised learning (WSL). To test our method, we collected H&E stained colorectal/stomach cancer slides (n=1,619) from multi-institutions and multiethnic groups, including TCGA Colorectal (CRC; n=351) and Stomach Adenocarcinoma (STAD; n=174) and six datasets from tertiary hospitals in Korea (i.e., Yonsei-1 (n=174), Yonsei-1-remade (n=146), Yonsei-2 (n=95), St. Mary-1 (n=48), St. Mary-2 (n=50), and Yonsei-CLASSIC trial (n=581)). We also compare our method to multiple deep learning methods, including Densenet, Resnet, Shufflenet, Googlenet, Squeezenet and two WSL models, MIL (Campanell et al., 2019) and CLAM (Lu et al., 2021). Result: In each cancer type, we train each model using TCGA dataset as discovery cohort and tested on the rest of datasets as validation cohort. We use the area under of roc curve (AUC) for the evaluation metric. For the CRC cohorts, our DGP method achieves overall best performance with 0.81 (Yonsei-1), 0.82 (Yonsei-1-remade), 0.89 (Yonsei-2), 0.85 (St. Mary-1) 0.75 (St. Mary-2) AUCs compared to the state-of-the art methods. For the STAD cohorts, our method also achieves overall best performance with 0.74 AUC (Yonsei-classic). In addition, incorporating the uncertainty in prediction measured by our method improves the model’s performance to predict MSI. Specifically, removing cases with prediction results with high uncertainty, which could lead false positive prediction, significantly improves MSI prediction performance. Taken together, these results demonstrate the robustness and generalisability of the DGP model for MSI prediction across multicenter and multiethnic datasets. Citation Format: Sunho Park, Hongming Xu, Sung Hak Lee, Jeonghyun Kang, Tae Hyun Hwang. Deep Gaussian process with uncertainty estimation Improves microsatellite instability prediction based on whole slide image: A retrospective multicenter and multiethnic cohort study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5012.