Predicting pathologic complete response in colorectal cancer patients after immunotherapy based on endoscopic biopsy and deep learning approach.

3532 Background: Immune checkpoint inhibitors (ICIs) have emerged as effective treatments for microsatellite instability-high (MSI-H)/deficient mismatch repair(dMMR) tumors in a select subset of colorectal cancer (CRC) patients. Patients sensitive to preoperative immunotherapy may have the opportunity to be exempted from surgery, while those insensitive may avoid unnecessary treatment. Tumor pathology provides rich biological insights. Several studies have indicated that deep learning algorithms can predict the efficacy of immunotherapy directly from digitized hematoxylin-eosin (H&E) stained Whole Slide Images (WSIs). However, their potential application in CRC immunotherapy remains underexplored. Based on WSIs of endoscopic biopsy, this study aims to construct a predictive model using deep learning approach to identify potential pathological complete response (pCR) in CRC patients after preoperative immunotherapy. Methods: This study enrolled CRC patients who received preoperative immunotherapy at West China Hospital, Sichuan University. Stratified randomization based on pathological outcomes was performed, assigning enrolled patients to the training set (70%) and the validation set (30%). WSIs of endoscopic biopsy were used for analysis. A predictive model was developed based on the Swin Transformer architecture, integrating convolutional neural networks (CNNs) with a self-attention mechanism. Pre-trained weights were employed for feature extraction, and the CLAM (Clustering-constrained Attention Multiple Instance Learning) framework was utilized to optimize pathological image analysis. The model’s performance was assessed in the validation cohort using the Receiver Operating Characteristic Curve (ROC) and Area Under the Curve (AUC) was calculated. Attention-based visualization analysis was further performed to identify the top patches that contributes to the determination of tumor response to preoperative immunotherapy. Results: 96 CRC patients treated with preoperative immunotherapy were included, with 67 in the training set and 29 in the validation set. A total of 278,901 512×512-pixel patches were generated by preprocessing 144 WSIs. A predictive model were established based on the training set and verified in the validation set. The model achieved an AUC of 0.82. Attention-based visualization analysis recognized the top 5% patches contributing to the determination of tumor response to preoperative immunotherapy, with 62.66% identified as tumor tissues and 37.34% identified as non-tumor tissues. Conclusions: Endoscopic biopsy based deep learning model, with distinct attention to tumor and non-tumor regions, may provide a novel and effective tool for predicting pCR after preoperative immunotherapy in CRC patients.