Automatic segmentation of regions of interest in vaginal brachytherapy

Abstract

The postoperative endometrial carcinoma treatment often includes radiotherapy (external radiotherapy and/or vaginal brachytherapy) to prevent the reappearance of the tumour. This project aims to improve the efficiency of the vaginal brachytherapy treatment by developing an automatic segmentation algorithm capable of delineating both the clinical target volume and the organs at risk, reducing the time required by experts to exert such task. In this project, we develop an AI-based framework that uses a V-Net architecture at its core. To train and evaluate the model, we use retrospective CT images and corresponding manual delineations from patients treated in Hospital Clinic. The creation of the algorithm was achieved successfully, resulting in a completely functional creator of automatic segmentations. About its performance, the results were found satisfactory in the cases of the vagina, the rectum and the bladder, having acceptable discrepancies in the dosimetry output. On the other hand, the bowel and the sigma models would require further improvements since the segmentations obtained didn’t match the ground truth. Overall, the project represents a step forward in the application of artificial intelligence algorithms to radiotherapy related processes.