A comparative evaluation of Deep Learning models for tumor segmentation: identifying the optimal framework for clinical integration

Accurate tumor segmentation is crucial for effective radiotherapy planning and surgical applications through 3D models and patient-specific instruments. Manual delineation is labor-intensive and subject to inter-observer variability, driving the need for robust automated solutions. This study evaluates the performance of four Deep Learning frameworks —U-Net, SegResNet, UNETR, and SwinUNETR— implemented via the MONAI framework. A comprehensive dataset of 126 pelvic tumors was employed for training and validation. Quantitative metrics (Dice Similarity Coefficient, Hausdorff Distance and Average Surface Distance) and qualitative evaluation were used to assess segmentation performance. Our results indicate that, although each model shows potential, U-Net delivers superior segmentation accuracy and computational efficiency, making it the most viable candidate for integration into clinical workflows. These findings underscore the transformative role of Deep Learning-based segmentation in enhancing treatment planning efficiency in clinical practice.