Deep learning-based mismatch repair prediction using colorectal cancer macroscopic images: a diagnostic study

Area of Science:

• Artificial Intelligence in Oncology
• Computational Pathology
• Medical Imaging Analysis

Background:

• Mismatch repair (MMR) testing is crucial for colorectal cancer (CRC) management but is resource-intensive.
• Current MMR assays require specialized facilities and significant time, limiting accessibility.
• There is a need for rapid, cost-effective screening tools for MMR status in CRC.

Purpose of the Study:

• To develop and validate a deep learning model for predicting MMR status in CRC patients using macroscopic images.
• To establish a rapid and cost-free screening method for MMR deficiency in colorectal cancer.
• To assess the explainability of the developed deep learning model.

Main Methods:

• A two-step deep learning approach was employed, utilizing DeepLabV3+ for lesion segmentation and Vision Transformer (ViT) for MMR classification.
• The study included 809 CRC patients, with macroscopic images captured immediately post-resection and MMR status confirmed via immunohistochemistry.
• Model performance was evaluated using Area Under the Curve (AUC), with explainability assessed through Gradient-weighted Class Activation Mapping (Grad-CAM) and Principal Component Analysis (PCA).

Main Results:

• The deep learning model achieved high predictive performance with an average AUC of 0.896 (internal) and 0.860 (independent) for MMR status.
• Excellent Negative Predictive Values (NPVs) of 0.987 (internal) and 0.978 (independent) were reported, indicating strong reliability.
• Explainability analyses (Grad-CAM, PCA) confirmed the model's transparency and interpretability.

Conclusions:

• A novel deep learning model accurately predicts MMR status from macroscopic colorectal cancer specimen images.
• This AI-driven approach demonstrates significant potential as a rapid and cost-effective screening tool for MMR deficiency in CRC.
• The model's explainability supports its clinical utility, particularly in time-sensitive scenarios.