Source-linked AI summary
UI-VISA: U-Net Initialized Vascular Image Segmentation Architecture
Asees Kaur, Suzanne S. Sindi, Erica M. Rutter
TL;DR
Vascular segmentation in DSA is difficult because thin, branching vessels challenge pixel-wise U-Net predictions and region growing depends on seed initialization. UI-VISA uses U-Net foreground predictions to initialize CNN-guided region growing, which refines the segmentation through local connectivity. Across 5-fold evaluation on 26 DSA images, UI-VISA had the most favorable reported performance balance, with significant clDice improvement but nonsignificant Dice improvement over U-Net.
Problem
Thin, elongated, branching vessels can produce fragmented U-Net predictions, while region growing is sensitive to seed initialization and computationally expensive.
Method
UI-VISA uses U-Net foreground predictions as informed seeds for a CNN-guided region-growing algorithm that iteratively refines vessel segmentation.
Results
UI-VISA achieves the most favorable balance among the three methods, with statistically significant clDice improvement over U-Net (p = 0.023) but nonsignificant Dice improvement (p = 0.104).
Takeaways & Limitations
UI-VISA preserves vascular connectivity better than U-Net alone and is more computationally efficient than standalone VISA.
Takeaways & Limitations
UI-VISA remains substantially slower than standalone U-Net inference, motivating further computational-cost reduction.
Abstract
from arXiv · showhide
Accurate segmentation of vascular structures in digital subtraction angiography (DSA) images remains challenging due to the thin, elongated, and branching nature of blood vessels. Pixel-wise deep learning approaches such as U-Net achieve strong general-purpose segmentation performance but often produce fragmented or discontinuous predictions in fine vascular regions, since they do not explicitly enforce structural connectivity. Region growing algorithms preserve spatial context and topological continuity, but are highly sensitive to seed point initialization and can be computationally expensive. We propose UI-VISA (U-Net Initialized Vascular Image Segmentation Architecture), a hybrid pipeline that combines the complementary strengths of both approaches. UI-VISA uses U-Net's foreground predictions as informed seed points for a CNN-guided region growing algorithm, which then iteratively refines the segmentation by enforcing local connectivity and recovering fine vessel details that U-Net alone tends to miss or over-predict. We evaluate UI-VISA against standalone U-Net and a prior region-growing-based method (VISA) using 5-fold cross-validation on 26 DSA images. UI-VISA achieves the highest mean Dice and clDice scores across folds, and a paired Wilcoxon signed-rank test shows the improvement in clDice is statistically significant ($p=0.023$), consistent with the method's design goal of preserving vascular connectivity, while the improvement in Dice does not reach significance ($p=0.104$).
1. Introduction
DSA supports clear visualization of cerebral vessels but manual tracing is laborious and variable. UI-VISA combines U-Net predictions with connectivity-aware region growing to refine vascular segmentations.
- DSA is the standard-of-care imaging method for diagnosing and guiding treatment of cerebrovascular diseases because it clearly visualizes blood vessels.
- Manual vessel-boundary tracing is tedious, error prone, and susceptible to inter- and intra-observer variability.
- U-Net often produces patchy probability maps, discontinuous vessel segments, and isolated artifacts in fine, elongated, branching vascular structures.
- Region growing conditions pixel inclusion on connectivity to the growing region, enforcing the topological continuity expected of vascular structures.
- UI-VISA uses U-Net foreground predictions as informed seeds for region growing, refining patchy segmentations while recovering fine structural details.
2. Related Work
Prior DSA segmentation evolved from manual and handcrafted methods toward deep learning, while region growing addresses the structural inconsistencies of independent pixel classification. CNN-guided region growing further replaces handcrafted similarity criteria with learned features.
- DSA offers high spatial and temporal resolution and is considered the gold standard for diagnosing cerebrovascular diseases.
- Handcrafted multi-scale Hessian enhancement can improve vessel visualization but may introduce blurring at vessel edges.
- U-Net is widely used for DSA vessel segmentation, including approaches pretrained on unlabeled angiograms before fine-tuning on labeled images.
- Independent pixel classification can yield locally plausible but globally inconsistent regions that fail to respect physical and biological contiguity.
- CNN-based region growing replaces handcrafted similarity criteria with learned features for deciding which neighboring pixels to add.
3. Methods
The method trains a patch-based U-Net and a CNN-guided region-growing system, then initializes growth from U-Net foreground predictions. Evaluation uses Dice and clDice to assess overlap and vascular connectivity.
- 3. Methods: The dataset contains 26 patient-specific 512 × 512 DSA images with manually traced masks, evaluated using 5-fold cross-validation.
- 3.2. Modified U-Net: The modified U-Net processes randomly sampled 128 × 128 patches and produces a single-channel vessel probability map through an encoder-decoder with skip connections.
- 3.3. VISA: VISA trains a CNN on 80 × 80 image patches to predict 3 × 3 foreground masks, then iteratively grows contiguous regions from seed pixels.
- 3.4. UI-VISA: UI-VISA replaces VISA’s 500 random seeds with thresholded U-Net foreground predictions, while retaining the subsequent region-growing procedure.
- 3.4. UI-VISA: The final UI-VISA segmentation is produced by region growing, which expands around U-Net seeds to recover fine details and enforce local connectivity.
- 3.5. Evaluation Metrics: Dice measures overlap, whereas clDice evaluates preservation of vessel continuity and topology using segmentation and ground-truth skeletons.
4. Results
Across qualitative examples and five-fold cross-validation, UI-VISA generally outperformed VISA and U-Net, while trading some inference speed for improved vascular-structure preservation. Statistical testing found a significant clDice improvement but not a significant Dice improvement.
- Qualitative comparison: UI-VISA produced coherent segmentations when U-Net predictions were fragmented, but U-Net achieved the best Dice in one case with UI-VISA adding skull false positives.The three test scenarios also included an image where all methods performed similarly.
- Cross-validation: UI-VISA achieved the highest Dice and clDice scores in four of five cross-validation folds and was highest overall across all images.In Fold 4, U-Net achieved the highest scores, but UI-VISA still outperformed VISA.
- Inference time: U-Net completed inference in under 40 seconds per fold, while UI-VISA was slower but 3–5 fold faster than VISA.The reported speed difference reflects the computational cost of iterative refinement relative to a single U-Net forward pass.
- Statistical comparison: 18 of 26 images favored UI-VISA for clDice, compared with 17 of 26 for Dice.These counts summarize per-image comparisons against U-Net.
- Statistical comparison: The clDice improvement over U-Net was significant at p = 0.023, whereas the Dice improvement was not significant at p = 0.104.The analysis used a one-sided Wilcoxon signed-rank test after both metric differences failed the Shapiro-Wilk normality test.
5. Discussion and Future Work
UI-VISA combines U-Net predictions with VISA’s region growing algorithm to improve seed initialization and produce accurate, connected vessel segmentations. It offers a favorable accuracy–connectivity–efficiency balance, but remains slower than standalone U-Net and requires further computational optimization.
- 5. Discussion and Future Work: UI-VISA combines U-Net predictions with VISA’s region growing algorithm to improve seed initialization and produce more accurate, connected vessel segmentations.Seeds are drawn from U-Net foreground predictions rather than selected randomly.
- 5. Discussion and Future Work: UI-VISA inference takes approximately 592–1812 seconds per fold, versus 2223–9341 seconds for standalone VISA, a roughly 3–5-fold reduction.The reported reduction results from informed seed initialization.
- 5. Discussion and Future Work: UI-VISA is more accurate and better connected than U-Net alone while being more computationally efficient than standalone VISA.The method consistently preserves vascular connectivity while reducing inference time relative to standalone VISA.
- 5. Discussion and Future Work: UI-VISA remains substantially slower than standalone U-Net inference despite computational savings from informed seed initialization.Future work proposes boundary-aware seeds and early termination to reduce RGA iterations.