Technical Summary: HTC-SGA Former for Coronary DSA Vessel Segmentation

1. Problem Statement

Accurate segmentation of coronary vessels in Digital Subtraction Angiography (DSA) is critical for the computer-aided diagnosis and treatment planning of coronary artery disease (CAD). However, existing methods face significant challenges due to the specific characteristics of DSA images:

• Thin, Low-Contrast Vessels: Distal branches often exhibit weak signals and low contrast against the background, making them difficult to detect.
• Severe Class Imbalance: The ratio of vessel pixels to background pixels is extremely skewed, leading models to bias toward background prediction.
• Global-Local Trade-off: Traditional Convolutional Neural Networks (CNNs) capture local features well but struggle with long-range anatomical context, while pure Transformers often lose fine-grained spatial details and suffer from high computational costs.
• Boundary Ambiguity: Motion artifacts and overlapping anatomical structures result in fragmented vessel continuity and imprecise boundary localization.

Current state-of-the-art methods often fail to simultaneously balance global contextual reasoning with the preservation of weak distal vessels and fine boundaries.

2. Methodology: HTC-SGA Former

The authors propose HTC-SGA Former, a lightweight hybrid framework that integrates a CNN encoder with a Transformer decoder. The architecture is designed to preserve local vessel morphology while modeling long-range dependencies.

2.1. Architecture Overview

• CNN Encoder: A lightweight encoder extracts hierarchical features using four convolutional stages. It employs 3×3 convolutions, batch normalization, ReLU, and dropout, followed by max-pooling for down-sampling. A bottleneck block expands the channel dimension to enrich high-level semantic representations before decoding.
• Transformer Decoder: Unlike standard hybrid models that use a Transformer bottleneck, HTC-SGA Former places the Transformer within the decoder to progressively enhance global-local contextual reasoning.

2.2. Key Components

1. Multi-Scale Global-Local Window Attention (MS-GLWA):

   • This block addresses the trade-off between global context and local detail.
   • Local Branch: Uses separable convolutions (depthwise 3×3 and pointwise 1×1) to preserve fine vessel morphology and boundary details.
   • Global Branch: Captures anatomical context via window-based multi-head self-attention. Unlike standard Swin Transformers, it generates Query, Key, and Value (QKV) tensors from multi-scale feature representations (using parallel 1×1, 3×3, 5×5, and depthwise convolutions).
   • Adaptive Scaling: A learnable temperature parameter ($\tau$) is introduced to adaptively scale attention logits, improving robustness in low-contrast and noisy images.
   • Fusion: Global and local features are fused via element-wise summation.

2. Self-Guided Feature Attention (SGFA):

   • Designed to refine weak vessel responses and suppress background interference.
   • Preliminary Mask Generation: Generates a coarse vessel probability map from decoder features to serve as a structural prior.
   • Residual Attention: Uses the mask to create attention weights that enhance vessel-consistent features via a residual formulation ($\mathbf{F}_{attended} = \mathbf{F} \odot (1 + \mathbf{A})$), avoiding the information loss of hard gating.
   • Multi-Scale Context: Employs parallel dilated convolutions (rates 1, 2, 4) to capture vascular features at varying scales before fusing them for final refinement.

3. Boundary-Weighted Adaptive Compound Loss (BWACL):

   • A novel loss function designed to handle severe class imbalance and improve thin-vessel boundary preservation.
   • Boundary Weighting: Generates a pixel-wise weight map that assigns higher weights (up to 1.5×) to thin-vessel boundaries identified via Euclidean distance transforms.
   • Compound Formulation: Combines a Weighted Focal Loss (to emphasize hard pixels) and a Weighted Dice Loss (to optimize overlap).
   • Adaptive Balancing: Introduces a dynamic balancing factor ($\beta$) that adjusts the contribution of the Focal and Dice terms based on real-time error metrics (Recall error vs. Boundary false-positive error). If vessel recovery is poor, the Focal term is weighted higher; if boundary precision is lacking, the Dice term is prioritized.

3. Key Contributions

The paper outlines five primary contributions:

1. HTC-SGA Former Framework: A lightweight hybrid CNN-Transformer architecture specifically for coronary DSA segmentation, combining a CNN encoder for local morphology and a Transformer decoder for global-local reasoning.
2. MS-GLWA Block: A novel attention mechanism integrating multi-scale window-based self-attention with separable convolutions to capture long-range dependencies while preserving fine vessel structures.
3. SGFA Module: A vessel-focused refinement module that uses self-guided attention and multi-scale context to enhance weak vessel responses and suppress background noise.
4. BWACL: A boundary-weighted adaptive compound loss that addresses class imbalance by emphasizing thin-vessel boundaries and dynamically balancing vessel recovery with boundary refinement.
5. Comprehensive Validation: Extensive experiments demonstrating the framework's superiority over 14 state-of-the-art methods and the cross-backbone applicability of the proposed loss function.

4. Experimental Results

The method was evaluated on a private dataset comprising 300 coronary DSA sequences (150 right and 150 left coronary artery images) from 50 patients, using 5-fold cross-validation.

• Performance vs. State-of-the-Art: HTC-SGA Former outperformed 14 competing methods (including U-Net variants, pure Transformers like SwinUNet, and hybrid models) on both right and left coronary subsets.
  • Right Subset: Achieved 88.58% Recall and 88.01% Dice, with an Average Surface Distance (ASD) of 0.6944.
  • Left Subset: Achieved 85.23% Recall and 82.89% Dice, with an ASD of 1.0256.
  • Efficiency: The model maintains a compact architecture with only 0.81M parameters, significantly fewer than many competitors (e.g., SwinUNet has 85.9M parameters).
• Loss Function Efficacy: When applied to four standard backbones (U-Net, Attention U-Net, U-Net++, Attention U-Net++), BWACL consistently improved Recall and boundary metrics (ASD/ACD) compared to standard BCE+Dice losses.
• Ablation Study:
  • Adding MS-GLWA components (Local Branch, Global Branch, Enhanced QKV) progressively improved performance.
  • The SGFA module provided significant gains in boundary accuracy (ASD/ACD reduction).
  • The full BWACL loss yielded the highest Recall and Dice scores, confirming its necessity for optimizing thin-vessel recovery.
• Qualitative Analysis: Visual comparisons showed that HTC-SGA Former better preserved thin distal branches, maintained vascular continuity, and reduced false positives in low-contrast regions compared to other methods.

5. Significance and Claims

The authors claim that HTC-SGA Former provides a practical, parameter-efficient solution for coronary vessel analysis. By integrating complementary global-local contextual modeling with vessel-focused refinement and adaptive optimization, the framework addresses the specific limitations of current methods in handling weak distal vessels and severe class imbalance.

The paper asserts that the proposed approach enables reliable and computationally efficient coronary vessel analysis, supporting quantitative image analysis and future computer-assisted cardiovascular interventions. The authors emphasize that the combination of architectural design (MS-GLWA, SGFA) and optimization strategy (BWACL) is key to achieving high accuracy without the computational burden of large-scale pure Transformer models.

Future work, as stated by the authors, involves validating the framework on larger multi-center datasets, investigating real-time implementations, and extending the methodology to other vascular imaging tasks (retinal, cerebral, peripheral) and angiographic modalities.