VISCERA-SAM: Adapting Segment Anything for Multi-Visceral Fetal Abdominal Ultrasound Segmentation

Minh H. N. Le, Khoa D. Pham, Tuan Vinh, Thanh-Huy Nguyen, Han H. Huynh, Khanh T. Q. Le, Anh Mai Vu, Ha N. T. Luong, Ulas Bagci, Min Xu, Carl Yang, Phat Kim Huynh, Nguyen Quoc Khanh Le

Keywords: Fetal ultrasound, Abdominal organ segmentation, Foundation models, Segment Anything Model, Boundary-aware learning

Abstract:

Foundation models such as the Segment Anything Model (SAM) have recently been explored for medical image segmentation, but their performance on fetal ultrasound, characterised by speckle noise, low contrast, and weak anatomical boundaries, remains insufficiently studied. We introduce VISCERA-SAM, a boundary-aware adaptation of the ultrasound-specific SAM-style foundation model UltraSAM for multi-organ fetal abdominal segmentation (abdominal aorta artery, intrahepatic umbilical vein, stomach, and liver area). Using UltraSAM as initialization, VISCERA-SAM is fine-tuned to segment four clinically relevant fetal structures (aorta, umbilical vein, liver, stomach) from single-point prompts.
The approach enhances the baseline with (i) boundary-focused losses (Hausdorff or boundary loss) in addition to Dice–focal training, (ii) largest-connected-component filtering to suppress spurious regions, and (iii) geometry-preserving augmentations to improve robustness to venous shape variability. On our fetal abdominal dataset, the UltraSAM baseline achieves a mean Dice of 0.884, mIoU of 0.796, and mean Hausdorff distance of 12.1. VISCERA-SAM improves these results to a mean Dice of 0.910 and mIoU of 0.830, while reducing mean Hausdorff distance by approximately 30% to 8.44; segmentation mAP increases from 0.589 to 0.67. Performance gains are consistent across all four organs, with the largest boundary improvements for liver and vein. These findings indicate that ultrasound native SAM-style pretraining, combined with lightweight boundary-aware adaptation, can provide accurate and contour-faithful fetal abdominal segmentations suitable for downstream biometric analysis.

Primary Subject Area: Application: Radiology

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 189