- Main
Vascular Image Segmentation Architecture for 2D and 3D Biomedical Images
- Kaur, Asees
- Advisor(s): Rutter, Erica;
- Sindi, Suzanne
Abstract
Vascular image segmentation is critical to biomedical image analysis, yet remains challenging due to the complex, thin, branching nature of blood vessels. Accurate segmentation is essential for diagnosing cerebrovascular diseases, surgical planning, and disease monitoring. Most approaches make pixel-wise predictions, which often fail to preserve structural contiguity, producing fragmented vessel segments. Medical imaging datasets are also often small, making it difficult to train deep learning models that require large amounts of data. This thesis introduces three pipelines combining local patch-based prediction with a region growing algorithm to produce accurate, connectivity-preserving vascular segmentations under limited data availability.First, VISA (Vascular Image Segmentation Architecture), a two-phase pipeline combining local patch-based prediction with a Region Growing Algorithm (RGA), segments 2D Digital Subtraction Angiography (DSA) images while maintaining vascular connectivity. By extracting densely overlapping snaps from each image, VISA generates many training samples from a limited dataset. VISA demonstrates competitive performance against a U-Net baseline, with stronger Dice and clDice scores across most folds, indicating better preservation of topological structure, though at the cost of more computation time due to the iterative region growing process.Second, UI-VISA (UNet Initialized VISA) replaces VISA's random seed initialization with informed seeds from U-Net predictions, improving accuracy and efficiency. Since these seeds focus on foreground regions, the RGA starts from meaningful vessel areas rather than the entire image, reducing the iterations needed for convergence. UI-VISA outperforms both standalone VISA and U-Net in four of five folds in Dice and clDice, while reducing inference time three to five times compared to VISA.Third, 3D-VISA extends VISA to three-dimensional volumetric data, applying the same two-phase design to segment dense vascular structures in 3D biomedical images. By extracting overlapping 40x40x40 volumetric snaps centered on every voxel, 3D-VISA generates many training samples from a single annotated volume. As a proof-of-concept on a synthetic vascular dataset, 3D-VISA achieves near-perfect Dice, Jaccard, and clDice scores, demonstrating the potential for volumetric segmentation under limited data.Together, these three pipelines demonstrate that combining local patch-based prediction with region growing algorithms is a promising direction for accurate, connectivity-preserving vascular segmentation in both 2D and 3D medical images.