Abstract 2111: Transfer learning for tumor mutation burden prediction and spatial heterogeneity analysis from histopathology slides in bladder cancer

Abstract

Abstract Backgrounds: Tumor mutation burden (TMB) is an emerging potential genomic biomarker, which may help in identifying patients likely to benefit from immunotherapy. DNA sequencing has been used to determine the TMB, but its utility is limited due to time-consuming preparation and relative high cost. In addition, while tumor might carry a mixture of TMB high/low cells, bulk DNA sequencing is limited to provide a full spectrum of TMB heterogeneity landscape. On the other hand, recent studies have shown that deep learning on pathology slides can predict tumor molecular subtypes and genetic changes and provide landscape of spatial heterogeneity of tumor. In this work, we apply transfer learning for predicting bladder cancer patient TMB and the spectrum of spatial heterogeneity of TMB within tumor using H&E stained whole slide images (WSIs). Methods: We present a general computational and deep learning framework to extract histological image features in WSIs and predicts patient TMB level with the machine learning model. Our method mainly consists of four steps. (1) we trained the tumor detector based on convolutional neural networks (CNN) to detect bladder cancer regions in WSIs. (2) we applied affinity propagation clustering method to select a subset of representative tumor tiles. (3) transfer learning based on Xception model is applied to extract features from representative tumor tiles. (4) the WSI is characterized by integrating feature vectors of representative tumor tiles. The SVM classifier is then trained to predict patient with either low or high TMB level. Finally, the spatial heterogeneity analysis of TMB is performed and incorporated with predicted TMB status to identify patient subgroups with distinct survival outcomes. Results: Experiments have been performed on 253 different patient pathology slides selected from The Cancer Genome Atlas (TCGA) bladder cancer cohort. Evaluations show that the trained CNN-based tumor detector provides over 90% sensitivity and precision in detecting cancer regions from WSIs. The trained SVM classifier can provide over 0.75 AUC value (95% CI, 0.683-0.802) in distinguishing low and high TMB patients. Using the spatial heterogeneity analysis of TMB status, we found that bladder cancer patients with more homogeneous TMB high status across tumor regions show significantly better survival outcomes compared to other patients carrying mixtures of TMB high/low status across tumor regions (log rank test p<0.05). Conclusion: Our study indicates that TMB levels and the spatial heterogeneity of TMB levels within tumors could be predicted using digitized H&E stained WSIs for bladder cancer patients. To the best of our knowledge, this is the first work to predict TMB status and its spatial heterogeneity to stratify patients in bladder cancer. Our method is extensible to histopathology images of other organs for predicting TMB or other clinical outcomes implying the broad and economical aspects of clinical utility of TMB prediction using WSIs. Citation Format: Hongming Xu, Sung Hak Lee, Tae Hyun Hwang. Transfer learning for tumor mutation burden prediction and spatial heterogeneity analysis from histopathology slides in bladder cancer [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2111.