Abstract
INTRODUCTION: Accurate automated segmentation of epistaxis (nosebleeds) from endoscopic images is critical for clinical diagnosis but is significantly hampered by the scarcity of annotated data and the inherent difficulty of precise lesion delineation. These challenges are particularly pronounced in resource-constrained healthcare environments, creating a pressing need for data-efficient deep learning solutions. METHODS: To address these limitations, we developed LESS-Net, a lightweight, semi-supervised segmentation framework. LESS-Net is designed to effectively leverage unlabeled data through a novel combination of consistency regularization and contrastive learning, which mitigates data distribution mismatches and class imbalance. The architecture incorporates an efficient MobileViT backbone and introduces a multi-scale feature fusion module to enhance segmentation accuracy beyond what is achievable with traditional skip-connections. RESULTS: Evaluated on a public Nasal Bleeding dataset, LESS-Net significantly outperformed seven state-of-the-art models. With only 50% of the data labeled, our model achieved a mean Intersection over Union (mIoU) of 82.51%, a Dice coefficient of 75.62%, and a mean Recall of 92.12%, while concurrently reducing model parameters by 73.8%. Notably, this semi-supervised performance surpassed that of all competitor models trained with 100% labeled data. The framework's robustness was further validated at extremely low label ratios of 25% and 5%. CONCLUSION: Ablation studies confirmed the distinct contribution of each architectural component to the model's overall efficacy. LESS-Net provides a powerful and data-efficient framework for medical image segmentation. Its demonstrated ability to achieve superior performance with limited supervision highlights its substantial potential to enhance AI-driven diagnostic capabilities and improve patient care in real-world clinical workflows, especially in underserved settings.