tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Attentive Temporal Fusion Network (ATFNet) for Multi-frame Coronary Vessel Segmentation in X-ray Angiography

Authors :

Pouya Babaei¹ Farshad Almasganj²

1- Department of Biomedical Engineering, Amirkabir University of Technology, Tehran, Iran 2- Department of Biomedical Engineering, Amirkabir University of Technology, Tehran, Iran

Keywords :

Attentive Temporal Fusion Network،Coronary vessel segmentation،X-ray coronary angiography،Spatial Attention Temporal Squeeze،Structured sparsity loss

Abstract :

X-ray coronary angiography remains the clinical gold standard for visualizing coronary lumen but presents major challenges for automated analysis: low vessel contrast, overlapping anatomy, catheter occlusion, breathing/heartbeat motion and extremely thin branching vessels that fracture easily in segmentation maps. To address these issues we propose ATFNet (Attentive Temporal Fusion Network), a compact UNet++–inspired architecture that ingests short temporal stacks (four successive frames) and fuses motion and appearance cues into a single 2-D prediction. Key components are (i) SATS (Spatial Attention Temporal Squeeze), a per-frame directional spatial attention and learned temporal fusion that compresses four frames into a channel-recalibrated 2-D representation; (ii) SE_ResBlock3D/2D units that provide residual learning with squeeze-and-excitation attention in the 3D encoder and 2D decoder; (iii) DSF (Deep Supervision Fusion), which combines coarse (spatial merge) and attentive (channel-reweighted) fine kernels from multiple decoder depths into one robust output; and (iv) a topology-aware StructuredSparsityLoss (BCE–Dice base + multi-scale tree norm) together with the Lion optimizer and scheduler to stabilise and accelerate training on modest clinical data. On a manually annotated clinical XCA set, ATFNet produces noticeably more continuous, less fragmented vessel masks and improved temporal stability compared with single-frame baselines; ablation studies confirm that SATS, DSF, SE-Res blocks and the Lion optimizer each contribute to the observed gains. These results indicate that compact, attention-augmented temporal fusion, combined with a tree-aware loss, can substantially improve coronary vessel continuity and segmentation fidelity in angiographic sequences.