A deep learning model based on Mamba for automatic segmentation in cervical cancer brachytherapy

Abstract

Abstract This study developed and evaluated an automatic segmentation model based on the Mamba framework (AM-UNet) for rapid and precise delineation of high-risk clinical target volume (HRCTV) and organs at risk (OARs) in cervical cancer brachytherapy. Using 694 CT scans from 179 cervical cancer patients, the performance of five models (AM-UNet, UNet, DeepLab V3, UNETR and nnU-Net) was compared. The models were assessed using the Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and dose-volume index (DVI). AM-UNet achieved mean DSCs of 0.862, 0.937, 0.823, and 0.725 for HRCTV, bladder, rectum, and sigmoid, respectively. Subjective evaluations showed 93.07% of AM-UNet predicted HRCTV were rated as clinically acceptable or needing minor adjustments, with no unacceptable cases. Dosimetric differences between AM-UNet-generated and manually delineated contours were within 1%, highlighting its potential for improving clinical workflows in brachytherapy.

Keywords

• Cervical cancer
• Brachytherapy
• Auto-segmentation
• Deep learning
• Computed