PanPrompt3D: A Prompt-Guided SAM-Based Framework for 3D Pancreatic Segmentation Across Multi-Sequence MRI.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, Huzhou Central Hospital, Fifth School of Clinical Medicine of Zhejiang Chinese Medical University, Huzhou, 313000, Zhejiang, China.
- Deepwise Artificial Intelligence Laboratory, Beijing, China.
- Department of Radiology, Huzhou Central Hospital, Affiliated Central Hospital Huzhou University, Huzhou, 313000, Zhejiang, China.
- Department of Hepatobiliary Surgery, Huzhou Central Hospital, Fifth School of Clinical Medicine of Zhejiang Chinese Medical University, Huzhou, 313000, Zhejiang, China.
- Department of Radiology, Huzhou Central Hospital, Fifth School of Clinical Medicine of Zhejiang Chinese Medical University, Huzhou, 313000, Zhejiang, China. [email protected].
- Precision Medicine Research Center, Huzhou Central Hospital, Fifth School of Clinical Medicine of Zhejiang Chinese Medical University, Huzhou, 313000, Zhejiang, China. [email protected].
- Department of Hepatobiliary Surgery, Huzhou Central Hospital, Fifth School of Clinical Medicine of Zhejiang Chinese Medical University, Huzhou, 313000, Zhejiang, China. [email protected].
- Huzhou Key Laboratory of Intelligent and Digital Precision Surgery, Huzhou, 313000, Zhejiang, China. [email protected].
- Boron Neutron Capture Cancer Treatment Technology, Zhejiang Engineering Research Center, Huzhou, 313000, Zhejiang, China. [email protected].
Abstract
Accurate three-dimensional pancreatic segmentation is important for pancreatic assessment, treatment planning, and quantitative follow-up. However, most existing segmentation methods have been developed primarily for computed tomography, and their performance across clinically used magnetic resonance imaging (MRI) sequences remains less well characterized. We propose PanPrompt3D, a prompt-guided 3D segmentation framework adapted from SAM-Med3D for pancreatic MRI. PanPrompt3D combines a hybrid Transformer-CNN encoder, 3D triplet attention, volumetric prompt encoding, a 3D mask decoder, and a volume-aware hybrid loss incorporating explicit volumetric consistency. The model was evaluated on public T1-weighted (T1W) and T2-weighted (T2W) MRI data and an in-house cohort of 285 water-sequence examinations. With five-point prompts, PanPrompt3D achieved Dice scores of 81.66%, 85.03%, and 85.07% on water-sequence, T1W, and T2W MRI, respectively. Compared with SAM-Med3D under the same prompting protocol, paired Dice improvements were 4.70, 5.80, and 4.99 percentage points, respectively, with all three comparisons remaining statistically significant after Holm correction. PanPrompt3D also achieved R<sup>2</sup> values above 0.9 between predicted and reference pancreatic volumes. Leave-one-sequence-out experiments showed reduced performance on unseen sequences, while external validation on CrossPan and AMOS demonstrated transfer to independent public MRI data, with a clear performance decrease on the unseen out-of-phase contrast. These findings support PanPrompt3D as a practical prompt-guided framework for multi-sequence pancreatic MRI segmentation while highlighting the remaining challenge of generalization to unseen contrasts.