Masked pre-training and pseudo-vessel enhancement-based semi-supervised cerebral vascular segmentation

Accurate segmentation of cerebral vessels in computed tomography angiography (CTA) and computed tomography perfusion (CTP) images is essential for diagnosing cerebrovascular diseases. However, due to the scarcity of annotated data and the complexity of small and branching vessels, fully supervised deep learning methods face limitations in clinical scenarios.

This study aims to develop a semi-supervised learning framework that can achieve high-accuracy cerebrovascular segmentation using only limited labeled data, while improving robustness to and enhancing small vessel continuity.

We propose a novel semi-supervised segmentation framework integrating three core components: (1) a dual-scale masked pre-training strategy, which jointly reconstructs masked vessel regions and performs segmentation to enhance vascular continuity; (2) pseudo-vessel augmentation, which simulates vessel-like artifacts to improve discrimination between true vessels and false positives; and (3) multi-level consistency losses, which enforce output consistency across different enhancement and model predictions to enhance learning under low supervision. The model was evaluated on 948 CTA and CTP scans using five annotation ratios (5%, 10%, 20%, 50%, and 100%).

With only 5% of labeled data, the proposed method achieved 98.7% of the performance of a fully supervised model. Under full supervision, it reached Dice scores of 0.8379 on CTP and 0.9177 on CTA, outperforming the second-best method by 1.12% and 0.87%, respectively.

The proposed semi-supervised framework enables accurate cerebrovascular segmentation with minimal labeled data. It effectively enhances small vessel continuity, reduces over-segmentation from noise, and demonstrates strong robustness and generalizability, making it a promising solution for real-world clinical applications.