LDFSAM: Localization Distillation-Enhanced Feature Prompting SAM for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Standard Segment Anything Model (SAM)-based methods in medical imaging use explicit geometric prompts like bounding boxes.
• These rigid constraints struggle with complex, deformable medical structures, leading to segmentation errors due to localization noise.

Purpose of the Study:

• To introduce a novel framework, Localization Distillation-Enhanced Feature Prompting SAM (LDFSAM), for improved medical image segmentation.
• To shift from discrete coordinate inputs to a latent feature prompting paradigm for more robust segmentation.

Main Methods:

• LDFSAM employs a lightweight prompt generator refined via Localization Distillation (LD).
• This generator injects multi-scale features into the SAM decoder as Dense Feature Prompts (DFPs) and Sparse Feature Prompts (SFPs).
• The method guides segmentation without relying on explicit box constraints.

Main Results:

• LDFSAM achieved Dice scores exceeding 0.91 on four public benchmarks (3D CBCT Tooth, ISIC 2018, MMOTU, Kvasir-SEG).
• The framework outperformed prior SAM-based baselines and conventional networks.
• Significant performance gains were observed in low-data regimes, demonstrating robust generalization on an in-house cohort.

Conclusions:

• LDFSAM provides a reliable solution for automated medical image segmentation by utilizing feature prompting.
• The method effectively overcomes limitations of rigid spatial constraints in standard SAM approaches.
• Its strong performance, particularly in data-scarce scenarios, highlights its potential for clinical applications.