Predicting colorectal cancer microsatellite instability with a self-attention-enabled convolutional neural network

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Genomic instability detection

Background:

• Microsatellite instability (MSI) is a crucial biomarker in colorectal cancer (CRC).
• Accurate MSI prediction from histopathology images can aid clinical decision-making.
• Current prediction methods require further optimization for improved accuracy and efficiency.

Purpose of the Study:

• To develop and validate a novel deep learning framework for predicting MSI in CRC.
• To integrate convolutional neural network (CNN) and self-attention models for enhanced feature extraction and aggregation.
• To assess the performance of the proposed method against established deep learning techniques.

Main Methods:

• A hybrid model combining INSIGHT (CNN) and WiseMSI (self-attention) was developed.
• Whole slide images from a multicenter Chinese CRC cohort were utilized.
• Tumor tiles were identified, features extracted using ResNet, and aggregated using attention-based pooling.
• Performance was evaluated using Area Under the Curve (AUC), specificity, and sensitivity.

Main Results:

• The INSIGHT model achieved an AUC of 0.985 for tumor patch classification.
• A high Spearman correlation (0.7909) was observed between pathologist and INSIGHT tumor cell fraction.
• WiseMSI demonstrated a specificity of 94.7%, sensitivity of 84.7%, and AUC of 0.954.
• The proposed method outperformed five other classic deep learning methods.

Conclusions:

• The combined INSIGHT and WiseMSI model offers a robust and accurate method for MSI prediction in CRC.
• This AI-driven approach holds potential for improving diagnostic accuracy and patient stratification.
• Further validation in diverse cohorts is warranted to confirm generalizability.