LESS-Net: a lightweight network for epistaxis image segmentation using similarity-based contrastive learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated segmentation of epistaxis (nosebleeds) from endoscopic images is crucial for diagnosis.
• Scarcity of annotated data and difficulty in lesion delineation hinder current methods, especially in resource-limited settings.

Purpose of the Study:

• To develop a data-efficient deep learning solution for epistaxis segmentation.
• To address limitations of data scarcity and improve diagnostic accuracy in medical imaging.

Main Methods:

• Developed LESS-Net, a lightweight, semi-supervised segmentation framework.
• Utilized consistency regularization and contrastive learning to leverage unlabeled data.
• Incorporated a MobileViT backbone and multi-scale feature fusion module.

Main Results:

• LESS-Net outperformed seven state-of-the-art models on a public Nasal Bleeding dataset.
• Achieved 82.51% mIoU and 75.62% Dice coefficient with only 50% labeled data, surpassing fully supervised models.
• Demonstrated robustness at extremely low label ratios (25% and 5%) and reduced model parameters by 73.8%.

Conclusions:

• LESS-Net offers a powerful and data-efficient approach to medical image segmentation.
• Its performance with limited supervision can enhance AI-driven diagnostics and patient care.
• The framework holds significant potential for real-world clinical workflows, particularly in underserved areas.