A multi-task deep learning framework for simultaneous prediction of microsatellite instability and tumor mutational burden in gastric cancer from histopathological images.

Background: The clinical management of gastric cancer (GC) increasingly relies on the biomarkers microsatellite instability (MSI) and tumor mutation burden (TMB) to identify patients likely to benefit from immunotherapy. However, their ubiquitous adoption is hampered by the high cost and complexity of next-generation sequencing. We hypothesized that a single deep learning model could simultaneously and accurately predict both biomarkers directly from routine histopathology slides, offering a transformative, cost-effective diagnostic tool. We aim to develop a multi-task deep learning framework to simultaneously predict MSI and TMB using routine histopathological images and clinical data. Methods: We presented a novel, interpretable, multi-task deep learning framework that concurrently predicted MSI and TMB status. Our model innovatively integrated whole slide images (WSIs) and clinical data in an end-to-end architecture. It employed a pre-trained ResNet50 for feature extraction, an attention mechanism to identify predictive image regions, and a Multimodal Compact Bilinear Pooling (MCBP) layer to fuse these image features with structured clinical data (gender, age, T/N/M stage). The model was trained on 312 patients from The Cancer Genome Atlas (TCGA). Furthermore, to ensure robustness, an expanded independent external validation cohort of 121 GC patients from our local center was incorporated from the Cancer Hospital, Chinese Academy of Medical Sciences. Results: The multimodal framework achieved robust performance in cross-validation, achieving area under the curve (AUC) values of 0.828 for MSI and 0.836 for TMB on the internal TCGA test set, outperforming standard models like ResNet18 and VGG. While the model achieved high AUCs internally, performance on the external validation set showed a moderate decrease due to domain shifts, yielding an AUC of 0.78 for MSI and 0.74 for TMB. Model interpretability was achieved through attention heatmaps, which revealed a significant spatial concordance between regions predictive of MSI and TMB from Quantitative spatial analysis, providing novel biological insight and validating our multi-task design. Conclusion: This work establishes the feasibility and accuracy of a unified, multi-task deep learning framework for the concurrent prediction of key immunotherapy biomarkers in gastric cancer. By leveraging routinely available histopathological images and clinical data, our method represents a significant innovation with immediate potential to lower the barrier to precision oncology in clinical practice. Our framework provides a cost-effective, preliminary screening tool for MSI and TMB in GC. Although external validation highlights challenges in generalizability across different scanners, this approach shows promise in triaging patients for immunotherapy.