An interpretable digital pathology framework for microsatellite instability prediction in colorectal cancer using an attention-enhanced Xception network
Yanzong Lin, Yan Huang, Chenjun Zhang, Zhaohui Liu, Yimin Li, Yunxia HuangObjective
To develop and evaluate an attention-enhanced Xception network for predicting microsatellite instability (MSI) status in colorectal cancer using routine hematoxylin and eosin-stained histopathological images.
Methods
We developed an Xception model with a Convolutional Block Attention Module (CBAM). The model was trained, validated, and tested on a large-scale public dataset containing 193,312 image patches derived from FFPE histology slides of CRC patients in the TCGA cohort. Data were partitioned at the patient level into training, validation, and test sets to prevent data leakage. Model performance was evaluated at both the patch and patient levels using accuracy, precision, recall, F1-score, and AUC, with 95% confidence intervals. Grad-CAM was employed to visualize model attention.
Results
The model achieved a patch-level AUC of 0.9551 and accuracy of 88.4%. At the patient level, AUC was 0.962 and accuracy was 90.6%. Grad-CAM visualization demonstrated spatially localized attention patterns within representative histopathological patches, providing qualitative insight into the regions contributing to model predictions.
Conclusion
The attention-enhanced Xception model provides a computationally efficient and interpretable solution for MSI status prediction from H&E-stained slides. It holds promise as a computational decision-support tool to complement traditional molecular assays, especially in resource-constrained environments, and supports the growing role of deep learning in precision oncology.