StructSAM: structure-aware prompt adaptation for robust lung cancer lesion segmentation in CT

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Accurate lung lesion segmentation in CT scans is vital but challenging due to image limitations.
• Foundation models like SAM struggle with medical images lacking 3D context and clear boundaries.

Purpose of the Study:

• To develop a robust volumetric segmentation framework for medical images, focusing on lung cancer detection.
• To adapt foundation models for improved performance in clinical settings by integrating structural information.

Main Methods:

• Introduced StructSAM, a structure-aware prompt adaptation framework.
• Incorporated anatomical priors and a 3D inter-slice aggregator for volumetric consistency.
• Utilized Parameter-Efficient Fine-Tuning (PEFT) for scalability.

Main Results:

• StructSAM achieved state-of-the-art lung nodule segmentation accuracy on the LIDC-IDRI dataset.
• Demonstrated robust generalization to other organs (KiTS19, MSD Pancreas) despite domain shifts.
• Outperformed classical methods and existing SAM adaptations.

Conclusions:

• Embedding structural priors into foundation models is a promising strategy for medical image segmentation.
• StructSAM offers a clinically reliable and efficient solution for volumetric segmentation.
• The framework shows potential for broad applicability across different anatomical structures.