AutoPromptSeg: Automated Decoupling of Uncertainty Prompts with SAM for semi-supervised medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised learning for disease diagnosis is hindered by limited annotated medical data.
• Semi-supervised learning (SSL) leverages unlabeled data to improve segmentation accuracy.
• Promptable foundation models offer new avenues, but require specific prompt annotations often absent in medical datasets.

Purpose of the Study:

• To introduce AutoPromptSeg, a novel semi-supervised 3D medical image segmentation method.
• To address the challenge of limited prompt annotations for foundation models in medical imaging.
• To enhance segmentation performance in data-scarce scenarios.

Main Methods:

• Developed the Decoupled Uncertainty Prompt Generator (DUPG) to create effective prompts.
• Employed the Channel Alignment and Fusion Architecture (CAFA) to align and enhance feature representation from unlabeled data.
• Utilized a semi-supervised learning framework combining prompt generation and feature alignment.

Main Results:

• Achieved state-of-the-art performance on Amos 2022, LA, and BraTS 2020 datasets.
• Demonstrated high segmentation accuracy with only 10% labeled data: 68.78% Dice on Amos 2022, 90.02% on LA, and 86.63% on BraTS 2020.
• Validated the effectiveness of AutoPromptSeg in data-scarce 3D medical image segmentation.

Conclusions:

• AutoPromptSeg offers a powerful solution for 3D medical image segmentation with limited labeled data.
• The integration of prompt generation and feature alignment significantly boosts performance.
• This framework shows great potential for improving AI-driven disease diagnosis in resource-constrained settings.