UA-UNet: Uncertainty Aware Pseudo-Label Generation in Residual U-Net for Medical Image Segmentation

Medical image segmentation is crucial for accurate diagnosis and treatment planning. Most segmentation models are developed through supervised learning, which requires access to large amounts of annotated data. However, acquiring such datasets is often expensive and time-consuming. Semi-supervised learning (SSL) approaches alleviate this challenge by leveraging both labeled and unlabeled data to improve model performance. While SSL methods have achieved promising results, they still have some limitations. Predictions made by these methods disregard the model uncertainty, leading to unreliable predictions. The predictions made by these models are often affected by the noise introduced during the SSL process. Additionally, these models fail to adapt well to complex medical images and are unable to scale well in out-of-domain samples. To address these limitations, we propose UA-UNet, an uncertainty-guided teacher-student model architecture built on top of the Residual U-Net for multi-class medical image segmentation. The model incorporates uncertainty estimation, which guides the generation of high-quality pseudo-labels from an ensemble of teacher models. By combining consistency regularization and pseudo-labeling, our method effectively reduces the influence of high-uncertainty regions while enhancing segmentation accuracy. We compared the model with 10 other methods on the 2023 Kidney Tumor Segmentation Challenge dataset (KiTS23). The proposed approach outperformed state-of-the-art models with a Dice score of 0.901 and IoU of 0.891. The proposed model also provides uncertainty maps, which could enhance the interpretability of the segmentation result. These features make UA-UNet a robust method for semi-supervised segmentation in medical imaging, particularly when labeled data is scarce.