Purpose: To develop and validate a prediction model for evaluating pathological complete response (pCR) to neoadjuvant chemoradiotherapy in patients with rectal cancer from rapid frozen sections.
Experimental Design: Because the prediction of pCR is related to tumor cells，The Cancer Genome Atlas Colon Adenocarcinoma database was enrolled in this study which was used to construct a tumor-positive region screening model by ResNet18. Then, after preprocessing and screening of tumor-positive regions, forty-eight rectal cancer patients (2013-2017) receiving neoadjuvant chemoradiotherapy were enrolled to construct a pCR prediction model based on ResNet50 with weakly supervised learning. And pCR prediction values of all the patients' rapid frozen sections were visualized by heat maps.
Results: For the tumor-positive region screening model, ten-fold cross-validation is used. And the best one has an accuracy rate of 94.1% in training set, 93.8% in validation set, and an AUC of 0.961 in validation. Meanwhile, for the pCR prediction model, a five-fold cross-validation was used. The average accuracy rate is 89.7% in training set, and the average accuracy rate is 78% in validation set.
Conclusion: It’s a breakthrough achievement that using deep learning to obtain the predictions of pCR from rapid frozen sections. The model with high prediction accuracy, strong generalization ability can assist the judgment and prediction of clinical work furtherly.
 

deep learning,rectal cancer,pathological complete response,rapid frozen sections