A lightweight dual-student mean teacher semi-supervised semantic segmentation method for skin lesions

Recent advancements in supervised models for skin lesion segmentation have demonstrated promising performance. However, their high computational demands and the expensive acquisition of pixel-level annotations present significant challenges in clinical applications. To address these issues, we propose a novel method named Efficient Group Enhanced Uncertainty-guided Dual-Student Mean Teacher (EGE-UDSMT), which incorporates three synergistic components to enhance model performance. First, introducing an additional student sub-network improves training stability by mitigating the risk of convergence to local optima. Second, the confidence maps generated by the sub-networks help produce high-quality pseudo-labels, thereby reducing model complexity and enhancing performance within the proposed semi-supervised learning framework. Lastly, customized lightweight modules further streamline the network architecture, leading to a significant reduction in computational complexity. As a result, our method achieves superior segmentation performance on the ISIC 2017 and ISIC 2018 datasets, reducing both parameter count and computational complexity by factors of 34 and 42 compared to UCMT (U-Net) while improving the Dice Similarity Coefficient (DSC) by 5.32 % and mean Intersection over Union (mIoU) by 0.96 %. The relevant code has been uploaded to https://github.com/SatineMilk/EGE-UDSMT.