Do Segmentation Models Understand Vascular Structure? A Blob-Based XAI Framework

TL;DR

This work addresses whether 3D vascular segmentation models leverage global anatomical context or rely primarily on local image cues. It introduces a graph-guided, blob-based XAI pipeline that attaches gradient-based attributions to anatomically meaningful Points of Interest (POIs) and analyzes Saliency blobs at multiple scales using a Frangi-based detector. Across two public vascular datasets, the study finds that attribution is overwhelmingly localized around POIs with weak correlation to vessel-level properties such as thickness, tubularity, or connectivity, indicating limited use of global vascular reasoning. The results highlight the need for structured explainability tools and suggest integrating topological priors or graph-based constraints to move segmentation models toward anatomically informed representations of vascular structures.

Abstract

Deep learning models have achieved impressive performance in medical image segmentation, yet their black-box nature limits clinical adoption. In vascular applications, trustworthy segmentation should rely on both local image cues and global anatomical structures, such as vessel connectivity or branching. However, the extent to which models leverage such global context remains unclear. We present a novel explainability pipeline for 3D vessel segmentation, combining gradient-based attribution with graph-guided point selection and a blob-based analysis of Saliency maps. Using vascular graphs extracted from ground truth, we define anatomically meaningful points of interest (POIs) and assess the contribution of input voxels via Saliency maps. These are analyzed at both global and local scales using a custom blob detector. Applied to IRCAD and Bullitt datasets, our analysis shows that model decisions are dominated by highly localized attribution blobs centered near POIs. Attribution features show little correlation with vessel-level properties such as thickness, tubularity, or connectivity -- suggesting limited use of global anatomical reasoning. Our results underline the importance of structured explainability tools and highlight the current limitations of segmentation models in capturing global vascular context.

Figures (8)

• Figure 1: The proposed XAI pipeline. The red curves represent the vascular graph, with the green point indicating the selected POI. The resulting Saliency map (orange) is processed to compute the blob mask (cyan) and analyzed using our blob-based approach. Finally, the Saliency characteristics are combined with the POI's vessel properties to assess their contribution to the model's decision.
• Figure 2: Images produced using https://www.voreen.uni-muenster.de. Vascular structure graphs from IRCAD (left) and Bullitt (right) samples. The vessel ground-truth is shown in grey, with cyan dots representing the graph nodes and red curves the indicating the centerlines.
• Figure 3: Pipeline to determine the vessel connectivity degree relative to the patch. (\ref{['fig:relative_degree_1']}) Patch extraction, which may distort the vessel structure. (\ref{['fig:relative_degree_2']}) Exclusion mask $M_{exc}$ with the POI shown as a green dot and vessel centerlines as red curves. (\ref{['fig:relative_degree_3']}) Resulting image $I_{cc}$ where only the vessels extending beyond $M_{exc}$ are retained to define relative connectivity.
• Figure 4: Blob detection on the Saliency map. (Left) Saliency map where the green dot represents the explained POI. (Middle) Frangi filter output. (Right) Blob mask of positive attribution.
• Figure 5: Violin plot of (a) the Fisher contrast-to-noise ratio and (b) the average L1-norm ratio between blob attribution and background attribution, depending on the dataset.