Visual In-Context Learning for Few-Shot Eczema Segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence in dermatology

Background:

• Automated eczema diagnosis from digital images aids patient self-monitoring.
• Eczema region segmentation is key for automated diagnosis.
• Current deep learning methods (CNN U-Net, Swin U-Net) require extensive annotated data.

Purpose of the Study:

• To investigate visual in-context learning for few-shot eczema segmentation.
• To evaluate the performance of SegGPT for eczema segmentation without model retraining.
• To explore the impact of prompt number on SegGPT performance.

Main Methods:

• Applied visual in-context learning using the generalist vision model SegGPT.
• Utilized a few representative images as prompts for SegGPT.
• Benchmarked SegGPT performance against a CNN U-Net trained on a large dataset.

Main Results:

• SegGPT with 2 prompts achieved higher mean Intersection over Union (mIoU: 36.69) than CNN U-Net trained on 428 images (mIoU: 32.60).
• Increasing the number of image prompts for SegGPT negatively impacted performance.
• Visual in-context learning demonstrated superior performance with minimal data.

Conclusions:

• Visual in-context learning offers a faster and more effective approach for eczema segmentation.
• SegGPT shows promise for developing inclusive dermatological solutions, especially for under-represented demographics.
• Few-shot learning strategies are crucial for overcoming data limitations in medical imaging AI.