Semi-supervised intracranial aneurysm segmentation via reliable weight selection

Early detection of intracranial aneurysms (IA) is crucial for risk reduction and treatment planning. Computed tomography angiography (CTA) has been widely applied in IA screening due to its rapid, cost-effective, and high-resolution capabilities. However, manual segmentation of IA from CTA images is labor-intensive and time-consuming, highlighting the urgent need for automated segmentation methods to improve diagnostic efficiency. This paper presents a reliable weight selection mean teacher framework for IA segmentation, leveraging semi-supervised learning to enhance network performance with limited labeled data. The framework consists of a student network and two teacher networks, while the teacher networks generate adversarial noise to guide the student network in learning. A novel reliable sample weight selection module employs ensemble learning to quantify prediction weights adaptively to minimize errors from individual models. Additionally, we integrate xLSTM module into the segmentation network to capture the contextual features between frames in the input data, while adversarial noise is introduced to bolster model robustness. Experiments show that our method improves Dice scores by at least 4.8% compared to the fully-supervised segmentation model baseline and at least 1.1% compared to other state-of-the-art semi-supervised methods under 20% labeled data. The source code and the dataset are available from https://github.com/jinxiaokuang/RWS-MT.