Effective contrast-enhanced preprocessing for intracranial artery segmentation in digital subtraction angiography

Objective.To develop and evaluate a contrast-enhanced preprocessing method for improving intracranial artery segmentation in digital subtraction angiography (DSA), with particular emphasis on preserving vascular connectivity (VC) in thin and branching vessels.Approach.DSA images were processed using three input conditions: no preprocessing, contrast-limited adaptive histogram equalization (CLAHE), and the proposed haze-inspired contrast-enhancement method incorporating transmission-map estimation, wavelet-based low-frequency suppression, and edge-preserving optimization. The resulting images were used as inputs to U-Net, U-Net++, and nnU-Net models. Segmentation performance was evaluated using conventional overlap metrics and VC, which was used to assess preservation of vessel continuity.Main results.The proposed preprocessing method produced modest improvements in conventional overlap-based metrics but substantially improved VC. In the U-Net++ model, the VC error decreased to 8.77 with the proposed method, compared with 23.26 without preprocessing and 19.15 with CLAHE. For the nnU-Net model, the proposed technique markedly reduced the VC value to 17.58 from 39.05 (None) and 30.92 (CLAHE). This represents reductions of approximately 55% and 43%, respectively. These findings indicate that the proposed preprocessing method is particularly effective for reducing vessel fragmentation and preserving thin vascular branches.Significance.The proposed preprocessing framework provides a model-agnostic strategy for improving topology-sensitive DSA vessel segmentation. Rather than primarily increasing global overlap scores, the method enhances vascular continuity, which is clinically relevant for cerebrovascular interpretation, interventional planning, and automated quantitative analysis.