Automatic Organ-at-Risk (OAR) Segmentation for Cervical Cancer Radiotherapy Using a U-Net Architecture with an Attention Gate Mechanism

Organ-at-risk (OAR) delineation in cervical cancer radiotherapy is essential to protect healthy tissues from excessive radiation exposure. However, manual delineation is time-consuming and prone to inter-observer and intra-observer variability, highlighting the need for efficient and consistent automated segmentation. This study compares standard 2D U-Net and 2D U-Net with an Attention Gate Mechanism for segmenting the bladder, rectum, bowel bag, and femoral heads using pelvic CT images from 100 cervical cancer patients at RSUD Dr. H. Abdul Moeloek, Lampung. The models were trained using a combination of Soft Dice Loss and Binary Cross Entropy and evaluated through patient-based 5-fold cross-validation. The Attention U-Net achieved a Dice Similarity Coefficient (DSC) of 0,8772 for the femoral heads. The Wilcoxon Signed-Rank Test (α = 0.05) revealed significant differences only in selected metrics for the femoral heads. In contrast, bowel bag segmentation performance declined, with a DSC reduction of 0,0203 and an increase in Hausdorff Distance 95 (HD95) of 2,37 mm. These findings indicate that the Attention Gate does not consistently improve segmentation performance and remains suboptimal for organs with substantial anatomical deformation and heterogeneity.