Contrastive Pre-Training and Multiple Instance Learning for Predicting Tumor Microsatellite Instability

Accurate classification between tumor MicroSatellite Stability (MSS) and Instability (MSI) is crucial in gastrointestinal (GI) cancer prognosis and treatment. In this paper, we present a novel two-stage weakly supervised methodology, leveraging the synergy of Multiple Instance Learning (MIL) and a unique Contrastive Clustering Network (CCNet), aimed at enhancing MSI prediction in Whole Slide Image (WSI) analysis of GI cancers. In our framework, we utilize a contrastive learning-based feature extractor, coupled with MIL's efficient labeling. Our approach shows notable improvement in MSI classification, outperforming existing methods. Experiments using colorectal cancer and stomach adenocarcinoma datasets demonstrate the model's efficacy and generalizability, marking an advance in computational pathology and cancer diagnostics. Furthermore, we explored the efficacy of transfer learning using our model, examining the performance of pretrained feature extractors from ImageNet and STAD datasets. Our framework outperforms existing methods when pretrained on STAD and transferred to CRC data.Clinical relevance- The potential to enhance the accuracy of gastrointestinal cancer diagnosis and prognosis through advanced machine learning techniques. By leveraging transfer learning and weakly supervised frameworks clinicians can benefit from improved MSI prediction in histopathological images, aiding in personalized treatment strategies and patient outcomes.