SLIM-UAM - Semi-Supervised Learning In Medical Imaging: Uncertainty-Attention Model

Deep CNN/ViT supervised models for medical imaging typically require substantial amounts of high-quality labeled data, which can be challenging in resource-poor settings. This paper proposes a novel attention-integrated uncertainty-aware semi-supervised framework for medical image segmentation. By implementing the mean-teacher model, a knowledge distillation technique, our framework leverages unlabeled data by encouraging consistent predictions under various perturbations. Enhanced with an attention mechanism, our model (SLIM-UAM) significantly improves segmentation performance tested on two datasets (Left Atrium Segmentation Dataset and Medical Segmentation Decathlon - Spleen Dataset). It not only outperforms SOTA semi-supervised methods but provides performance similar to SOTA supervised model trained using 4x more labeled data. Our model uses semi-supervised learning for medical image segmentation, targeting the problem of expensive labelling by medical experts.