Uncertainty-Guided Anatomical Prompt Refinement of Medical-SAM3 for Cross-Domain Pancreatic Tumor Segmentation in Three-Dimensional Computed Tomography
Year: 2026
Type: article
Abstract: Pancreatic tumors are small, low in contrast, and change appearance across scanners and centers, so their CT segmentation degrades across centers. Promptable foundation models such as Medical-SAM3 depend on prompt quality and give no signal of where a segmentation is unreliable. UAPR-SAM3 keeps Medical-SAM3 frozen and optimizes its prompts using prior knowledge and predictive uncertainty. An anatomical prompt generator derives prompts from an organ-relative coordinate frame, a tumor location atlas, and a hypoattenuation prior; an uncertainty estimation module separates epistemic from aleatoric uncertainty; a refinement loop turns uncertain regions into new prompts until the uncertainty mass stops decreasing; and a fusion stage returns a calibrated three-dimensional mask with an uncertainty map. On the MSD Task07 test split, UAPR-SAM3 reaches 62.4% tumor Dice, exceeding DTS-Gen by 5.1 tumor Dice points, PanSegNet by 1.4 pancreas Dice points, and a LoRA fine-tuned Medical-SAM3 by 3.5 tumor Dice points. On a fully independent PANORAMA subset (578 PDAC cases from two centers not used in development), it reaches 54.7% tumor Dice, which is 6.9 points above the strongest baseline. On 80 tumor-free TCIA Pancreas-CT scans, pancreas Dice is 83.9%, and a false-positive tumor occurs in 5 scans (0.09 mL per scan), without any weight update.
Source: Electronics
Institutions: Islamia University of Bahawalpur, Michigan State University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225756305