SimICL: A Simple Visual In-context Learning Framework for Ultrasound Segmentation.

Area of Science:

• Computer Vision
• Medical Imaging AI
• Self-Supervised Learning

Background:

• Conventional deep learning models require extensive expert labeling for medical imaging tasks, limiting generalizability.
• Visual in-context learning (ICL) offers a promising alternative by enabling models to adapt to new tasks with few examples.
• Current ICL methods face challenges in efficiency and performance, especially with limited annotated data.

Purpose of the Study:

• To introduce SimICL, a simple yet effective visual in-context learning method.
• To evaluate SimICL's performance in segmenting bony structures in wrist ultrasound (US) images.
• To demonstrate the potential of SimICL in reducing manual annotation efforts for AI model training in medical imaging.

Main Methods:

• SimICL combines visual in-context learning with masked image modeling (MIM) for self-supervised learning.
• The method was validated on a wrist ultrasound dataset for bony structure segmentation.
• A test set of 3822 images from 18 patients was used for evaluation.

Main Results:

• SimICL achieved a Dice coefficient (DC) of 0.96 and Jaccard Index (IoU) of 0.92 for bony region segmentation.
• These results significantly surpass state-of-the-art segmentation and visual ICL models, with improvements of at least 0.10 (DC) and 0.16 (IoU).
• The high performance was achieved with limited manual annotations.

Conclusions:

• SimICL demonstrates remarkable efficacy in segmenting bony structures in ultrasound images, even with limited data.
• The method significantly reduces the need for expert manual labeling, decreasing human expert time.
• SimICL enhances the real-world applicability of AI assistance in ultrasound image analysis.