Uncertainty-aware semi-supervised learning for enhanced multi-organ segmentation in CT volumes

We propose an uncertainty-aware semi-supervised learning (SSL) strategy for multi-organ segmentation in CT volumes. CT volume segmentation is crucial for enhancing diagnostic accuracy, facilitating quantitative analysis, and enabling automation and assisted diagnosis. Previous SSL strategies have shown great performance on CT volumes by ensuring consistency between different outputs of the model. However, these methods do not address the noise and uncertainty within individual outputs, which can impact the results. Our method integrates uncertainty estimation into the SSL framework, dynamically generating masks based on the uncertainty of individual predictions. This approach not only maintains consistency between multiple outputs but also mitigates the impact of noise and uncertainty within each prediction. Under the 10% labeled data setting, comparative experiments demonstrated that our method improved segmentation performance across effectively segmented organs. Specifically, compared to the baseline, the Dice scores increased by 2.50% for the vein, 1.99% for the liver, 5.16% for the spleen, and 6.09% for the stomach, indicating the potential benefits of incorporating uncertainty estimation into the SSL framework.