A Cascade-Based Model for Cerebral Artery Segmentation Using CASED-Net

Abstract

Cerebral artery segmentation of Digital Subtraction Angiography (DSA) images is an essential and fundamental step in computer-aided diagnosis (CAD) and treatment planning for cerebrovascular diseases such as intracranial aneurysm, arteriovenous malformation (AVM) and ischemic stroke. Although great progress has been made in deep learning, thin and meandering vascular structures, significant vessel size variations, low-contrast areas and background interference make accurate segmentation of cerebral vessels difficult. Current convolutional neural network (CNN)-based methods usually have poor receptive fields, lack of multi-scale feature representation, shallow-deep feature fusion and redundant feature propagation, which will lead to continuous vessel segmentation and lack of structural consistency. In order to solve these problems, a novel dual-view Encoder-Decoder multi-source learning network named CASED-Net (Cascade-Aware Encoder-Decoder multi-source learning network) for accurate segmentation of cerebral artery from DSA images is proposed. The proposed framework offers a combination of Anterior-Posterior (AP) and Lateral DSA projections to be used for learning complementary vascular representations. Fine-grained multi-scale features are captured by using a Res2Net-50 backbone, and the proposed Multi-source Feature Optimization Module (MFOM) increases cross-view feature extraction and adaptive feature refinement. A Multiscale Adaptive Context Module (MACM) is proposed to effectively combine contextual information from multiple feature scales, followed by a Cascaded Dilated Decoding Module (CDDM) that preserves vessels continuity by incrementally expanded receptive fields. Moreover, a Multi-directional Feature Enhancement (MFE) Module is designed to enhance deep-shallow feature interactions and boundary refinements; the Redundant Feature Suppression Module (RFSM) suppresses redundant responses via cosine similarity-based feature alignment, which further improves the consistency of structures and segmentation accuracy. The proposed CASED-Net was tested on the publicly available DIAS dataset, and was compared to several leading methods for cerebral vessel segmentation. The proposed framework is shown to perform well in segmentation with a Dice Similarity Coefficient of 0.8512, Accuracies of 97.11%, Sensitivities of 81.12%, Specificities of 99.47%, IoU of 0.7845 and AUC of 0.9744, while maintaining fine vascular structures and enhancing the connectivity of vessels. The findings show that CASED-Net is a powerful and reliable tool for automatic segmentation of cerebral artery and has great potential in computer-aided diagnosis and clinical decision-making in neurovascular imaging.