FNPC-SAM: Uncertainty-Guided False Negative/Positive Control for SAM on Noisy Medical Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• The Segment Anything Model (SAM) is a powerful foundation model for general image segmentation.
• SAM exhibits limitations in segmenting medical images, particularly low-contrast and noisy ultrasound data.
• Existing methods require extensive training or fine-tuning for medical image segmentation.

Purpose of the Study:

• To improve SAM's performance and robustness for medical image segmentation, especially in challenging ultrasound datasets.
• To introduce a test-phase prompt augmentation technique that enhances SAM's segmentation accuracy without retraining.
• To enable 3D segmentation from single 2D slices using a novel method.

Main Methods:

• Developed a prompt augmentation technique combining multi-box prompts with an aleatoric uncertainty-based false-negative (FN) and false-positive (FP) correction (FNPC) strategy.
• Evaluated the proposed method on two distinct ultrasound datasets.
• Introduced the Single-Slice-to-Volume (SS2V) method for 3D segmentation using single 2D slice annotations.

Main Results:

• Demonstrated significant improvement in SAM's segmentation performance on ultrasound images.
• Showcased enhanced robustness to inaccurate prompts without requiring additional training or model tuning.
• Successfully enabled 3D segmentation from single 2D slices, increasing efficiency.

Conclusions:

• The proposed prompt augmentation and FNPC strategy effectively enhances SAM for medical image segmentation, particularly in noisy, low-contrast conditions.
• The method offers a practical solution for leveraging SAM in medical imaging without the need for extensive retraining.
• The SS2V method provides an efficient approach for 3D segmentation tasks using limited annotations.