VessShape: Few-shot 2D blood vessel segmentation by leveraging shape priors from synthetic images

TL;DR

The paper tackles the data scarcity and cross-domain generalization challenges in blood vessel segmentation by introducing VessShape, a synthetic 2D data generator that enforces a strong vessel-shape bias through Bézier-curve–based geometries while varying textures. By pre-training U-Net backbones on VessShape and then fine-tuning on real datasets (DRIVE and VessMAP), the authors demonstrate improved few-shot segmentation performance and notable zero-shot capabilities across domains. The approach yields rapid convergence and robustness to domain shifts, indicating that shape priors can be a powerful, data-efficient alternative to texture-focused learning. This work suggests practical improvements for vascular segmentation in settings with limited annotated data and varying imaging modalities, and points to extensions into 3D and other tubular biological structures.

Abstract

Semantic segmentation of blood vessels is an important task in medical image analysis, but its progress is often hindered by the scarcity of large annotated datasets and the poor generalization of models across different imaging modalities. A key aspect is the tendency of Convolutional Neural Networks (CNNs) to learn texture-based features, which limits their performance when applied to new domains with different visual characteristics. We hypothesize that leveraging geometric priors of vessel shapes, such as their tubular and branching nature, can lead to more robust and data-efficient models. To investigate this, we introduce VessShape, a methodology for generating large-scale 2D synthetic datasets designed to instill a shape bias in segmentation models. VessShape images contain procedurally generated tubular geometries combined with a wide variety of foreground and background textures, encouraging models to learn shape cues rather than textures. We demonstrate that a model pre-trained on VessShape images achieves strong few-shot segmentation performance on two real-world datasets from different domains, requiring only four to ten samples for fine-tuning. Furthermore, the model exhibits notable zero-shot capabilities, effectively segmenting vessels in unseen domains without any target-specific training. Our results indicate that pre-training with a strong shape bias can be an effective strategy to overcome data scarcity and improve model generalization in blood vessel segmentation.

Figures (5)

• Figure 1: Illustration of the universality of vessel shape. (a) Fluorescence microscopy sample of a mouse cortex. (b) Fundus photograph of a human eye. Despite having different textures, the vessels have similar shapes.
• Figure 2: Examples of the VessShape generation process. Textures are sampled from ImageNet and blended according to procedurally generated binary masks to create the final VessShape images.
• Figure 3: Samples from the DRIVE and VessMAP datasets and their respective ground-truth masks.
• Figure 4: Few-shot and zero-shot Dice performance on (a) DRIVE and (b) VessMAP. Curves show the mean Dice over $R{=}5$ runs and $S{=}3$ repetitions for each sample size $n$. The shaded areas represent the standard deviation among runs. The inset shows the zero-shot Dice ($n{=}0$) for VSUNet50 on DRIVE, which was much lower than on the other experimental conditions.
• Figure 5: Qualitative visual comparison of segmentation performance on the DRIVE and VessMAP datasets. The figure shows the outputs of VSUNet and U-Net variants across different few-shot regimes: zero-shot, one-shot, and few-shot (16-shot for DRIVE and 20-shot for VessMAP). For the U-Net variants, no zero-shot output is available. Red squares highlight specific regions of interest to facilitate comparison: an area with low-caliber vessels in DRIVE and a vascular bulge in VessMAP.