Date of Publication : June 2026

Abstract: Accurate segmentation of brain tumours from magnetic resonance imaging (MRI) scans is critical for diagnosis, treatment planning, and monitoring of neurological diseases. This paper presents a comparative study between two fundamentally different segmentation paradigms: (1) a task-specific U-Net model trained from scratch on brain MRI data, and (2) Meta’s Segment Anything Model (SAM), a large-scale foundation model applied in a zero-shot setting with bounding box prompts. Using the LGG (Lower-Grade Glioma) MRI Segmentation dataset, we evaluate both approaches on an identical validation set of 786 slices (266 tumour-containing) using five metrics: Dice coefficient, IoU, 95th-percentile Hausdorff distance (HD95), sensitivity, and precision. SAM achieves superior overall segmentation quality (Dice: 0.877 vs. 0.745, p < 0.001) and significantly better boundary precision (HD95: 5.37 vs. 15.53 pixels) compared to U-Net. On tumour-containing slices, both models achieve comparable Dice scores (0.637 vs. 0.638) but exhibit fundamentally different error profiles: SAM favours high sensitivity (0.987) at the cost of precision (0.494), while U-Net favours precision (0.750) over sensitivity (0.661). These findings demonstrate that foundation models hold significant promise for medical image segmentation in interactive clinical workflows, while task-specific models remain the practical choice for fully automated pipelines.

Reference :