A Large-Scale Morphologically Labeled Dataset and Validation Framework for 2D Interventional Abdominal Angiography
Abstract
Purpose
AI-based medical image segmentation has achieved strong performance across diverse tasks, yet progress remains constrained by ground truth availability. Interventional fluoroscopic angiography, a 2D modality with complex vascular anatomy, is underrepresented in public repositories. This study presents the Multi-Label Interventional Vascular Dataset (HFH.MIVD), largest publicly available annotated dataset of 2D abdominal interventional angiograms (n=1005), featuring innovative morphological labels and an iterative validation methodology applicable to future dataset construction.
Methods
Retrospective review enrolled 246 patients comprising 5436 celiac axis angiograms, filtered to 1108 candidate images after quality analysis. Vascular structures were manually annotated using three morphological classes defined by vessel diameter ratio relative to catheter insertion diameter: Trunk (>45%), Bifurcation (20-45% arising from Trunk or Bifurcation vessels), and Periphery (≤20%). These thresholds stratify segmentation complexity by distinguishing high-contrast central vessels from increasingly challenging peripheral structures. A multi-stage validation process included blinded review, physician adjudication, and interrater analysis.
Results
Adjudication yielded 1005/1108 accepted images; exclusions resulted from contrast extravasation obscuring anatomy, motion-induced artifact, and vessel ambiguity. Cohen's Kappa demonstrated strong label agreement (μ=0.827, σ=0.08). Hausdorff Distance-95 showed robust structural agreement between annotations and validations (μ=4.7 pixels, σ=6.16 pixels) with low inter-labeler disparity (μmax=11.50 pixels, σmax=2.11 pixels). Convexity analysis validated the morphological classification scheme, with non-overlapping interquartile ranges across classes: 0.458 [IQR 0.299-0.707], 0.158 [IQR 0.111-0.222], and 0.018 [IQR 0.010-0.041] for Trunk, Bifurcation, and Periphery respectively.
Conclusion
HFH.MIVD addresses a gap in publicly available vascular segmentation resources for abdominal interventional imaging. The morphological labeling scheme enables class-specific model training targeting distinct segmentation difficulty levels. The iterative validation pipeline provides a reproducible framework for constructing large, annotated datasets involving complex anatomical targets and/or alternative labeling schema.