Unveiling the Clinical Significance of Microsatellite Instability in Colorectal Cancer: Deep Learning and the Segment Anything Model for Accurate Segmentation and Classification

Abstract

Microsatellite instability (MSI) is crucial for colorectal cancer (CRC) diagnosis and prognosis. Accurate differentiation between MSI and microsatellite stability (MSS) tumors is essential for personalized treatment. This paper introduces a novel approach combining the segment anything model (SAM), Yolov8, and convolutional neural networks (CNNs) for precise segmentation and classification of histopathological images. SAM employs a prompt-based mechanism for segmenting tumor regions like invasive margins, tumor-infiltrating lymphocytes (TILs), and necrotic areas. Integrating SAM’s segmentation with CNN-based classification achieves high-accuracy MSI-H/MSS subtyping by focusing on key histopathological features. Tested on TCGA-CRC data, this approach outperformed traditional methods in segmentation and classification accuracy, enhancing MSI/MSS diagnostic potential and enabling efficient high-throughput analysis in clinical and research settings.