Endoscopy-Based Deep Convolutional Neural Network Predicts Response to Neoadjuvant Treatment for Locally Advanced Rectal Cancer.

Abstract

Background and Aims: Although the wait and watch (W&W) strategy is a treatment choice for locally advanced rectal cancer (LARC) patients who achieve clinical complete response (cCR) after neoadjuvant therapy (NT), the issue on consistency between cCR and pathological CR (pCR) remains unsettled. Herein, we aimed to develop a deep convolutional neural network (DCNN) model using endoscopic images of LARC patients after NT to distinguish tumor regression grade (TRG) 0 from non-TRG0, thus providing strength in identifying surgery candidates. Methods: A total of 1000 LARC patients (6,939 endoscopic images) who underwent radical surgery after NT from April 2013 to April 2021 at the Sixth Affiliated Hospital, Sun Yat-sen University were retrospectively included in our study. Patients were divided into three cohorts in chronological order: the training set for constructing the model, the validation set, and the independent test set for validating its predictive capability. Besides, we compared the model’s performance with that of three endoscopists on a class-balanced, randomly selected subset of 20 patients’ LARC images (10 TRG0 patients with 70 images and 10 non-TRG0 patients with 72 images). The measures used to evaluate the efficacy for identifying TRG0 included overall accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUROC). Results: There were 219 (21.9%) cases of TRG0 in the included patients. The constructed DCNN model in the training set obtained an excellent performance with good accuracy of 94.21%, specificity of 94.39%, NPV of 98.11%, and AUROC of 0.94. The validation set showed accuracy, specificity, NPV, and AUROC of 92.13%, 93.04%, 96.69%, and 0.95, respectively; the corresponding values in the independent set were 87.14%, 92.98%, 91.37%, and 0.77, respectively. In the reader study, the model outperformed the three experienced endoscopists with an AUROC of 0.85. Conclusions: The proposed DCNN model achieved high specificity and NPV in detecting TRG0 LARC tumors after NT, with a better performance than experienced endoscopists. As a supplement to radiological images, this model may serve as a useful tool for identifying surgery candidates in LARC patients after NT.