TL;DR
This paper introduces BAVT, a Boundary-Aware Vision Transformer, which enhances standard Vision Transformers (ViTs) with an edge-aware loss function. This novel approach significantly improves the segmentation of fine vascular structures in coronary angiography, outperforming existing CNN and hybrid methods while remaining compatible with large-scale pretraining.
What Problem Does This Paper Address?
Accurately segmenting vascular networks in medical images like coronary angiograms is crucial for diagnosing diseases. However, these vessels are often thin, elongated, and have low contrast, making them difficult to detect and delineate precisely. While traditional Convolutional Neural Networks (CNNs) struggle with maintaining the topological continuity of these structures, recent Vision Transformers (ViTs), despite their strength in capturing global context, often lack the fine-grained boundary awareness needed for accurate segmentation. This gap hinders their direct application in clinical settings where precise vessel mapping is essential.
Methods
The core innovation of this work is the Boundary-Aware Vision Transformer (BAVT). BAVT builds upon a standard ViT architecture, specifically an encoder-decoder setup. The key enhancement lies in the introduction of an edge-aware loss function during training. This loss explicitly penalizes errors at the boundaries of vascular structures, guiding the model to learn more precise segmentations. Unlike some hybrid approaches that combine CNNs and Transformers, BAVT maintains a pure ViT structure, ensuring compatibility with modern Vision Foundation Model (VFM) pretraining strategies like DINOv2 and EVA-02. The edge-aware loss is formulated using a signed distance map of the ground truth mask, encouraging high prediction confidence near true boundaries and penalizing false positives/negatives in thin regions.
Datasets
The primary dataset used for training and evaluation is the DCA-1 coronary angiography dataset. This dataset consists of 134 grayscale angiographic images with expert-annotated ground truth masks. The authors note that the DCA-1 dataset lacks an official benchmark split, so they adopted a standardized protocol, allocating 77.6% for training and 22.4% for testing, and performed experiments using four independent random splits to ensure robustness.
Key Insights and Ideas
- Edge-Awareness is Crucial for Fine Structures: The paper strongly suggests that for tasks involving thin, elongated structures like blood vessels, explicit boundary supervision is more critical than just global context modeling.
- ViT Compatibility Matters: By sticking to a plain ViT architecture, BAVT can readily leverage powerful pre-trained foundation models, offering a significant advantage for generalization on challenging medical datasets.
- Minimal Architectural Change, Maximum Impact: The introduction of a novel loss function, rather than complex architectural modifications, proves highly effective in improving segmentation quality.
- Curriculum Learning's Untapped Potential: The authors highlight that curriculum learning can significantly improve performance in segmenting peripheral and microvascular structures, an aspect often overlooked.
Key Results
BAVT achieved superior performance on the DCA-1 dataset across various medical image segmentation metrics, including Sensitivity, Specificity, F1-score, Accuracy, AUC, and IoU. Specifically, BAVT achieved an F1 score of 0.81, outperforming both CNN-based baselines (like U-Net) and hybrid transformer-CNN models. Qualitative results show BAVT producing more continuous and anatomically accurate vessel structures, particularly in thin branches, compared to other methods that tend to over-smooth or fragment these delicate features. The ablation study confirmed that the edge-aware block significantly improved sensitivity and IoU.
Limitations and Future Work
The authors acknowledge that while BAVT is computationally feasible, its complexity is still higher than traditional CNNs, estimated at approximately 130–140 GFLOPs per forward pass for a 512×512 input. Future work includes extending the architecture to handle 3D angiographic sequences and multi-view vessel tracking. They also plan to investigate integrating uncertainty estimation into the ViT pipeline to enhance clinical trustworthiness and usability.
Summary updated