Semantic Segmentation Refiner for Ultrasound Applications with Zero-Shot Foundation Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning shows promise in medical imaging but struggles with segmentation due to limited labeled data.
• A domain gap exists between natural and medical images, particularly ultrasound, hindering model fine-tuning.
• Segmentation model performance degrades significantly in low-data regimes.

Purpose of the Study:

• To address performance degradation in medical image segmentation within low-data environments.
• To propose a prompt-less segmentation method using foundation models for abstract shape segmentation.
• To evaluate the method on ultrasound image segmentation for pathologic anomalies.

Main Methods:

• Developed a novel prompt point generation algorithm using coarse semantic segmentation masks.
• Utilized a zero-shot, prompt-able foundation model as an optimization target.
• Applied the method to a segmentation findings task on ultrasound images.

Main Results:

• Demonstrated effectiveness on a musculoskeletal ultrasound dataset across varying low-data regimes.
• Achieved greater performance gains as the training dataset size decreased.
• Showcased advantages in segmenting pathologic anomalies in ultrasound images.

Conclusions:

• The proposed prompt-less method effectively enhances medical image segmentation in low-data scenarios.
• Foundation models can be harnessed for abstract shape segmentation, overcoming data scarcity challenges.
• The approach offers a promising solution for improving diagnostic accuracy in ultrasound imaging with limited data.