Development of deep learning-based automatic MRI segmentation for brachytherapy planning in cervical cancer

Abstract

Manual delineation of organs at risk (OARs) is an essential step in MRI-guided brachytherapy planning for cervical cancer. However, this process is time-consuming and may be affected by inter-observer variation. This study aimed to develop and evaluate a deep learning-based automatic segmentation model for the bladder, rectum, and sigmoid on T2-weighted MRI for cervical cancer brachytherapy planning. The model performance was evaluated using geometric metrics, including the Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95). In addition to geometric evaluation, a clinical assessment by radiation oncologists was conducted to determine the model's practical usability. The model achieved the highest performance for bladder segmentation, with a mean DSC of 0.92 ± 0.03 and a mean HD95 of 2.67 ± 1.16 mm. Rectum segmentation showed intermediate performance, with a mean DSC of 0.79 ± 0.05 and a mean HD95 of 7.32 ± 2.49 mm. Sigmoid segmentation was the most challenging, with a mean DSC of 0.66 ± 0.15 and a mean HD95 of 32.49 ± 27.27 mm. From the clinical evaluation, 90% of bladder contours and 52.5% of rectum contours were considered acceptable with only minor or no modifications, whereas sigmoid contours more frequently required substantial revision, indicating limited reliability for this structure. In conclusion, the proposed model has the potential to improve the efficiency and consistency of OARs contouring in MRI-guided cervical cancer brachytherapy planning, particularly for the bladder and rectum. The short inference time (less than 1 second per case) further supports its feasibility for routine clinical workflow. However, sigmoid auto-segmentation still requires further refinement to enhance segmentation robustness before routine clinical implementation.