A Deep Learning Classifier Based on Pre-Radiation Computed Tomography and Clinical Parameters to Predict Pathological Complete Response after Neoadjuvant Chemoradiation in Esophageal Cancer

Purpose/Objective(s)

Neoadjuvant chemoradiation (NCRT) followed by surgery is the standard treatment for resectable esophageal cancer. More than 30% patients achieve pathological complete response (pCR) after NCRT, who may avoid the followed surgery. However, there is no reliable method in predicting pCR yet. Artificial intelligence, especially deep learning, has made great progress in many fields including treatment response prediction. Therefore, we built up a deep learning classifier based on pre-radiation computed tomography and clinical parameters to predict pCR after NCRT for esophageal cancer.

Materials/Methods

Between 2009 and 2021, consecutive patients with esophageal cancer received NCRT and complete resection were retrospectively analyzed. Pathological response assessed on surgical specimen was collected. Patients were randomly assigned to the training set, validation set, and testing set as 7: 1: 2. We built a binary classification neural network based on 3D Resnet. Pre-radiation computed tomography (CT) was fed as input to build the imaging classifier. The filtered clinical parameters including gender, tumor location, clinical stage, pathological type, sequence of chemoradiation, chemotherapy regimen and radiotherapy technique were then added by encoded as fully connected layer to build the combined classifier. Area under the receiver operating characteristic curve (AUC) was calculated to evaluate the prediction performance and the optimal cut-off point was determined by Youden index.

Results

Totally 279 patients were enrolled, of whom 93 achieved pCR (33.3%). The performances of imaging classifier were AUC=0.989 (95% CI 0.937-0.986) with the sensitivity of 98.6% and specificity of 98.5% in the training set, and AUC=0.649 (95% CI 0.481-0.660) with the sensitivity of 66.7% and specificity of 58.9% in the testing set, respectively. After the addition of clinical parameters, the combined classifier showed AUC=0.855 (95% CI 0.797-0.986) with the sensitivity of 82.4% and specificity of 74.3% in the training set, and AUC=0.731 (95%CI 0.631-0.819) with the sensitivity of 76.6% and specificity of 65.6% in the testing set, respectively.

Conclusion

The combined deep learning classifier can accurately predict pCR after NCRT for esophageal cancer. Besides, addition of necessary clinical parameters can remedy the overfitting of imaging classifier. Prospective exploration based on larger data sets is needed to further improve the accuracy and generalization.