Image-Based Prediction of Microsatellite Instability in Colorectal Cancer Using H&E-Stained Histopathology: A Dual-Path Deep Learning Approach

Background and Aims: Microsatellite instability (MSI) is a key molecular biomarker in colorectal cancer, critical for identifying patients eligible for immunotherapy and screening hereditary cancer syndromes such as Lynch syndrome. Current gold-standard MSI detection methods—immunohistochemistry (IHC) and PCR—are accurate but often time-consuming, resource-intensive, and inaccessible in low-resource settings due to their dependence on specialized laboratory infrastructure and expertise. These constraints hinder routine MSI screening in many clinical environments. Recently, deep learning has emerged as a promising non-invasive alternative, with growing evidence supporting the feasibility of predicting MSI status from routine hematoxylin and eosin (H&E) stained slides. However, the lack of publicly available datasets with matched MSI labels and the homogeneity of prior studies limit model generalizability. This study aims to develop and validate a dual-path deep learning model for MSI prediction using a novel, diverse dataset of H&E-stained colorectal cancer images, offering a scalable, low-cost adjunct to molecular testing. Methods: Our proposed dual-path deep learning method integrates features from two pre-trained convolutional neural networks—DenseNet۱۲۱ and InceptionV۳—fine-tuned on a private dataset comprising over ۵۰۰ digitized H&E-stained colon cancer images with confirmed MSI status. We collected data from a referral pathology laboratory with diverse staining protocols and patient demographics. To enhance feature representation, we applied a Convolutional Block Attention Module (CBAM) to the outputs of each path. The attention-weighted features from both paths were globally pooled, concatenated, and passed through fully connected layers for final binary classification. Model performance was evaluated using cross-validation metrics such as accuracy, F۱-score, and AUC. Results: The final model achieved a training accuracy of ۹۸.۴% and an F۱-score of ۹۸.۳%. Validation accuracy and F۱-score reached ۹۵.۱% and ۹۴.۹%, respectively. On the test set, the model maintained robust performance, achieving ۹۶.۳% accuracy and ۹۶.۹% F۱-score, confirming its generalizability across heterogeneous data.

کلیدواژه ها:

Microsatellite instability ، Colorectal cancer ، Deep learning ، Histopathology ، H&E-stained slides