Evaluating deep learning models for Pelvic Bone Tumor Segmentation: Implications for Radiotherapy and Surgical Applications

Background: Accurate segmentation of pelvic bone tumors is crucial for effective treatment planning in radiotherapy and surgical interventions. Manual segmentation is labor-intensive and subject to variability, highlighting the need for automated solutions. This study evaluated the performance of four deep learning (DL) frameworks- U-Net, SegResNet, UNETR, and SwinUNETR-in automating the segmentation of pelvic bone tumors in a high- Complexity hospital setting, aiming to identify the most viable options for clinical integration. Methods: A cohort of 78 patients with pelvic bone tumors from a tertiary care hospital, including patients aged 14-88 years, was used. The dataset underwent preprocessing, involving DICOM to NIfTI format conversion and focused cropping on tumor regions. These data were then divided into training, validation, and test sets. Each DL framework was trained on the same pre-processed data, with variations in hyperparameters such as image size, batch size, and data augmentation, to optimize performance. The models were evaluated based on the Dice similarity coefficient (DSC), 95% Hausdorff distance (95% HD), and average surface distance (ASD), along with training time and qualitative visual assessment. Results: Among the four frameworks, U-Net and SwinUNETR demonstrated the best balance between segmentation accuracy and computational efficiency. U-Net achieved a DSC of (81.79 ± 21.84)% with training times of 15 minutes and 36 seconds, making it particularly suitable for environments with limited computational resources. SwinUNETR, despite longer training times, delivered the highest DSC of (82.08 ± 0.23)%. Visual evaluations confirmed that SwinUNETR and UNETR indeed provided the most visually accurate segmentations, closely aligning with the ground truth. Conclusions: U-Net and SwinUNETR are identified as the most clinically viable DL frameworks for pelvic bone tumor segmentation, offering an optimal balance between accuracy, Computational efficiency and resource demands. Despite limitations in GPU memory and dataset size, this study contributes to the integration of automated segmentation into clinical workflows. These findings provide a strong foundation for further optimization of these models and their scalability across different tumor types, aiming to enhance patient care in oncology and improve medical imaging practices.