Uncertainty Guided Joint Cross Pseudo Supervision for Semi-Supervised Intracranial Aneurysm Segmentation

Early detection of intracranial aneurysms (IA) is crucial for preventing rupture and guiding clinical intervention. Although computed tomography angiography (CTA) has been widely used in IA diagnosis due to its clear imaging and ease of use, manually locating aneurysms in CTA images is time-consuming and subjective, significantly impacting efficiency. Deep learning models have recently demonstrated strong performance in image segmentation tasks, but their reliance on large amounts of labeled data has become a bottleneck for clinical application. To address this problem, we propose a simple yet effective semi-supervised learning approach, called uncertainty guided joint cross pseudo supervision (UJ-CPS). Our basic framework consists of a shared encoder and three decoders with identical architectures but independent parameters. First, we propose an uncertainty-based joint pseudo-label generation mechanism, which weights the outputs of two decoders based on uncertainty to produce pseudo-labels with higher confidence. Second, we use the joint pseudo-labels to supervise the output of the rest decoder. The cross-supervision mechanism mitigates the error accumulation problem caused by low-quality pseudo-labels. Furthermore, because learning a more general and richer feature representation from three different views, our shared encoder approach improves the encoder's feature extraction capabilities. Finally, during testing, we let the three decoders jointly predict by leveraging uncertainty to produce more accurate segmentation results. Experimental results on two datasets show that our method performs competitively with the fully-supervised segmentation model baselines and several state-of-the-art methods in terms of Dice coefficient, Jaccard index, sensitivity and 95% Hausdorff distance, validating its effectiveness and generalization capabilities.