Deep learning model for predicting treatment response for patients with rectal cancer using 3D medical imaging.

Abstract

168 Background: We developed a deep learning model to predict pathological response based on magnetic 3D resonance imaging in rectal cancer. Methods: A total of 242 patients with locally advanced rectal cancer who received preoperative chemoradiotherapy followed by surgical resection were collected from single center. Surgical resection was performed at 6 to 8 weeks after the completion of preoperative CRT. Achieving pathogic complete response (pCR) was defined as complete absence of any tumour cells, in both the primary site and the dissected lymph node in surgical specimens. Based on pre-chemoradiotherapy T1-weighted axial 3D MR images, deep learning models were developed to predict pCR, respectively. Results: We analyzed 32,000 MRI images and involved several radiation oncologists in segmenation and classification. To calculate the probability of pCR using various convolutional neural network (CNN) architectures, several deep learning models were developed. The data input to the deep learning model was 3D MRI image, clinical data, and MRI metadata. Among 242 patients, 50 (20.7%) had evidence of pCR. The best model was developed based on SeResNet using transfer learning by MONA framework. The deep learning model showed an area under the receiver operating characteristic curve, sensitivity, specificity, and accuracy of 0.926, 0.831, 0.921, and 90.2% for predicting pCR. Conclusions: The pre-chemoradiotherapy T1-weighted 3D MR image-based deep learning model showed acceptable performance in predicting pCR in patients with rectal cancer.[Table: see text]