Automated Segmentation of After-Loaded Metal Source Applicators in Cervical Cancer Treatment Using U-Net: Enhancing Efficiency and Accuracy in Treatment Planning

Abstract

This study utilized the U-Net deep learning model to automate the segmentation of three-dimensional after-loaded metal source applicators, aiming to expedite treatment planning, reduce patient wait times, and enhance the treatment process. Using CT images from cervical cancer patients treated between December 2020 and August 2023, 27 images formed the training set, 3 were for validation, and 10 for testing. The model’s performance was evaluated against expert delineations using metrics like the Dice similarity coefficient (DSC), Hausdorff distance 95% (HD95), and others. The results were integrated into an after-loading planning system to locate applicator pathways and assess dose accuracy and feasibility. For the test group, the DSC ranged from 0.90 to 0.93, HD95 from 0.79 to 0.80 mm, and ASSD from 0.03 to 0.22 mm, with an average segmentation time of 65 seconds, significantly faster than manual delineation. The automatic pathways closely matched the original plan’s dosimetric parameters (P >0.05), indicating the system’s potential for safe application in after-loading planning for cervical cancer treatment. The U-Net-based region-growing method shows promise in improving the efficiency and accuracy of after-loaded applicator segmentation.

Keywords

• Cervical cancer
• after-loaded treatment planning
• deep learning
• U-Net
• automatic segmentation