A prediction model for pathological findings after neoadjuvant chemoradiotherapy for resectable locally advanced esophageal squamous cell carcinoma based on endoscopic images using deep learning.

OBJECTIVES To propose deep learning (DL)-based predictive model for pathological complete response rate for resectable locally advanced esophageal squamous cell carcinoma (SCC) after neoadjuvant chemoradiotherapy (NCRT) with endoscopic images. METHODS AND MATERIAL This retrospective study analyzed 98 patients with locally advanced esophagus cancer treated by preoperative chemoradiotherapy followed by surgery from 2004 to 2016. The patient data were split into two sets: 72 patients for the training of models and 26 patients for testing of the model. Patients was classified into two groups with the LC (Group I: responder and Group II: non-responder). The scanned images were converted into joint photographic experts group (JPEG) format and resized to 150 × 150 pixels. The input image without imaging filter (w/o filter) and with Laplacian, Sobel, and wavelet imaging filters deep learning model to predict the pathological CR with a convolution neural network (CNN). The accuracy, sensitivity, and specificity, the area under the curve (AUC) of the receiver operating characteristic were evaluated. RESULTS The average of accuracy for the cross-validation was 0.64 for w/o filter, 0.69 for Laplacian filter, 0.71 for Sobel filter, and 0.81 for wavelet filter, respectively. The average of sensitivity for the cross-validation was 0.80 for w/o filter, 0.81 for Laplacian filter, 0.67 for Sobel filter, and 0.80 for wavelet filter, respectively. The average of specificity for the cross-validation was 0.37 for w/o filter, 0.55 for Laplacian filter, 0.68 for Sobel filter, and 0.81 for wavelet filter, respectively. From the ROC curve, the average AUC for the cross-validation was 0.58 for w/o filter, 0.67 for Laplacian filter, 0.73 for Sobel filter, and 0.83 for wavelet filter, respectively. CONCLUSIONS The current study proposed the improvement the accuracy of the DL-based prediction model with the imaging filters. With the imaging filters, the accuracy was significantly improved. The model can be supported to assist clinical oncologists to have a more accurate expectations of the treatment outcome. ADVANCES IN KNOWLEDGE The accuracy of the prediction for the local control after radiotherapy can improve with the input image with the imaging filter for deep learning.