Abstract
Natural images often contain curvilinear structures that may be disconnected or partly occluded. Recovering the missing connections of disconnected structures is an open issue and requires appropriate geometric reasoning. We propose to find line co-occurrence statistics from the centerlines of blood vessels in retinal images and show its remarkable similarity to a well-known probabilistic model for the connectivity pattern in the primary visual cortex. Furthermore, the probabilistic model is trained from the data via statistics and used for automated grouping of interrupted vessels in a spectral clustering-based approach. Several challenging image patches are investigated around junction points, where successful results indicate the perfect match of the trained model to the profiles of blood vessels in retinal images. Also, comparisons among several statistical models obtained from different datasets reveals their high similarity, indicating they are independent of the dataset. On top of that, the best approximation of the statistical model with the symmetrized extension of the probabilistic model on the projective line bundle is found with a least square error smaller than 2%. Apparently, the direction process on the projective line bundle is a good continuation model for vessels in retinal images.
Introduction
The accurate mapping of retinal vasculature is a cornerstone in diagnosing and monitoring systemic diseases such as hypertension and diabetic retinopathy. However, automated segmentation algorithms often struggle with "broken" data where vessels appear disconnected due to low contrast or ambiguity at bifurcations and crossings. While humans effortlessly perceive these as continuous contours—guided by innate Gestalt laws of perception—computational models traditionally rely on manually designed cost functions that often fail in complex scenarios. This research addresses the challenge of automated perceptual grouping by teaching computers to reason about vessel connectivity using the underlying geometric logic of vessel layouts.
Methodology and Data
The proposed framework bridges biological and computer vision through several key steps. First, 2D retinal images are "lifted" into a 3D orientation score space ($R^2 \times P^1$). This transformation allows the algorithm to distinguish between vessels that cross at different angles at the same spatial location. The study utilized the DRIVE and IOSTAR datasets, which provide high-resolution color fundus and Scanning Laser Ophthalmoscope (SLO) images, respectively.
Using manual ground-truth segmentations from these datasets, the authors calculated line co-occurrence statistics—the probability that two vessel segments at specific relative positions and orientations belong to the same structure. This data-driven kernel is then integrated into a self-tuning spectral clustering algorithm to group disconnected pixels into individual, continuous vessels without requiring manual thresholding.
Results and Discussion
A primary finding of this study is that the statistical distribution of vessel connections in the retina matches theoretical models of the human visual cortex (specifically Mumford’s direction process) with an error rate as low as 0.33% to 1.04%. This suggests that the "continuation logic" employed by the brain is remarkably well-suited for medical imaging tasks.
The results also highlight dataset independence; kernels derived from the DRIVE dataset were nearly identical to those from IOSTAR. This robustness implies that the geometric rules of vessel growth are universal, allowing models trained on one imaging modality to be applied to others. Qualitatively, the spectral clustering approach successfully resolved complex junctions and crossings, even in the absence of strong intensity information, indicating a perfect match between the trained model and retinal vessel profiles.
Conclusion
By modeling vessel connectivity on a projective line bundle, this research provides a robust mathematical framework for retrieving challenging vessel connections. The high degree of alignment between empirical data and biological vision models validates the use of stochastic direction processes for medical image analysis. Future work aims to incorporate curvature and scale into the statistical models to further improve grouping accuracy and to extend this approach to other curvilinear structures, such as corneal nerve fibers or road networks in satellite imagery.
Summary updated