Accurate delineation of organs at risk (OARs) in head and neck CT images is essential for safe and effective radiotherapy. While U-Net-based deep learning models perform well, segmenting small, complex OARs is still difficult due to low contrast and class imbalance. This study investigates the impact of a slice-based (cropping) preprocessing strategy on the segmentation accuracy, consistency, and computational efficiency of a Residual U-Net framework for head and neck OAR delineation. A total of 63 CT scans covering 41 OARs were used. Networks were trained separately for each organ with full images and organ-specific crops. Segmentation accuracy was assessed using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU). Paired Wilcoxon tests compared full-size and slice-based models. Additional experiments with dropout and extended training targeted challenging structures like the optic chiasm and optic nerves. Slice-based preprocessing improved IoU by 4.1% and Dice score by 3.2% across 41 OARs (p < 0.001). For 11 small, complex organs, gains were 10.9% in IoU and 9.0% in Dice (p < 0.001). It also reduced performance variability, indicating better consistency. Training time dropped to less than half, about 2.3 times faster, while inference speed increased eightfold. Dropout and extended training slightly improved optic pathway metrics, but not significantly. The proposed slice-based Residual U-Net framework improves segmentation accuracy and computational efficiency for head and neck OAR delineation. This approach is particularly beneficial for small and anatomically complex structures and may provide a practical solution for integration into radiotherapy planning workflows.

Keywords