FeTal-SAM: Atlas-prompted Segment Anything Model for fetal brain MRI achieves flexible segmentation without retraining

FeTal-SAM adapted SAM with atlas-based prompts for fetal brain MRI segmentation. On two datasets, Dice scores matched dataset-specific baselines for well-contrasted structures (cortical plate, cerebellum) but were lower for subtle structures like hippocampus. The method segments…

• The model uses multi-atlas registration to generate dense prompts, enabling binary per-structure segmentation that is fused into 3D volumes.
• Performance was comparable to state-of-the-art baselines trained for specific label definitions, but lower accuracy was observed for low-contrast structures (e.g., hippocampus, amygdala).
• The approach eliminates the need to retrain for new parcellations by conditioning on available atlas image–label pairs.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.