AISCT-SAM: Customized SAM-Med2D with 3D Context Awareness and Self-Prompt Generation for Fully Automatic Acute...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neurology and Stroke Research

Background:

• Accurate segmentation of acute ischemic stroke (AIS) lesions on Non-Contrast CT (NCCT) is vital for timely diagnosis and treatment.
• Existing semi-automatic models like SAM-Med2D struggle with performance, 3D context integration, and efficient prompting for AIS lesion segmentation on NCCT.

Purpose of the Study:

• To introduce AISCT-SAM, a novel SAM-Med2D-based model for automatic and accurate AIS lesion segmentation on NCCT scans.
• To enhance segmentation by incorporating 3D volumetric context and developing a self-prompt generation mechanism.

Main Methods:

• Proposed an adapter and low-rank adaptation with a gate layer for flexible fine-tuning to handle lesion heterogeneity.
• Introduced a plug-and-play depth adapter for extracting 3D contextual information and improving local priors.
• Developed a self-prompt generator leveraging bilateral hemisphere differences for fully automatic segmentation and a prompt-guided mask decoder tailored for NCCT characteristics.

Main Results:

• AISCT-SAM achieved Dice scores of 63.32%, 48.75%, and 43.81% on three diverse datasets, outperforming 20 state-of-the-art methods.
• Volumetric analysis and external validation indicate its utility in AIS diagnosis and treatment planning.
• Demonstrated good performance on a public MRI dataset, suggesting potential for cross-modality generalization.

Conclusions:

• AISCT-SAM offers a significant advancement in automatic AIS lesion segmentation on NCCT, addressing limitations of previous methods.
• The model's 3D context awareness and self-prompting capabilities enhance segmentation accuracy and efficiency.
• AISCT-SAM shows promise as a valuable tool for clinical decision-making in AIS and potential for broader medical imaging applications.