Deep-learning based rectal tumor segmentation on multi-parametric MRI using pre-treatment and near-end-treatment data

In this study, the convolutional neural network (CNN) was implemented to segment the rectal cancer in 185 patients, with tumor ROI outlined by a radiologist on multiparametric magnetic resonance images (mp-MRI). The Dice similarity coefficient (DSC) and mean symmetrical surface distance (MSD) values were used to evaluate the results of the proposed algorithm. The mean DSC and MSD were 0.88 and 2.2 cm respectively for pre-treatment MR images. The transfer-learning model using paired pre-treatment and near-end-treatment images could also lead to acceptable result for boost volume segmentation with the mean DSC of 0.83 and MSD of 2.7 cm. Our work showed the deep-learning with combined image sequences was promising for automatic tumor localization and segmentation of locally advanced rectal cancer. After applying transfer learning with the prior information obtained from the pre-treatment images, the proposed method can be further used to identify boost volume for radiation dose escalation which has been demonstrated beneficial in improving pathological complete response rate.