Precise Center Point Determination of Brain Vessel Branch Points and Endpoints After Deep Learning Object Detection

Abstract

Accurate localization of vascular landmarks, including branching points and endpoint(s), is fundamental for quantitative morphological analysis in neurovascular research and plays a critical role in the early detection of cerebrovascular pathologies, risk assessment, and treatment planning. Although deep learning-based object detectors have achieved high accuracy in identifying these landmarks, their bounding box outputs inherently lack the coordinate precision required for detailed morphometric studies such as three-dimensional vessel reconstruction and hemodynamic modeling. This study aims to address this limitation by proposing a novel hybrid methodology that integrates the robust detection capabilities of a deep learning model with the pixel-level precision of a targeted morphological post-processing stage. In the proposed approach, a Faster R-CNN model first identifies candidate regions by generating bounding boxes, and subsequently, a hit-or-miss transform (HMT) is applied within each bounding box to determine the precise center coordinates. The methodology was validated on 96 skeletonized slices from a single mouse brain—the High-Fat/High-Sucrose 13 (HFHS13) vasculature dataset, which exhibits significant topological heterogeneity—using a held-out test set (76 training / 20 test slices) with bootstrap 95% confidence intervals; because the slices are consecutive z-sections of one brain, we additionally verified performance under a leakage-free contiguous block split. We emphasize that the evaluation scope is limited to a single subject and that the method depends on high-quality upstream vessel segmentation and skeletonization. Experimental results demonstrate strong performance on the held-out test set, with precision rates of 99.86% for endpoint(s) and 99.82% for branching points, corresponding recall rates of 99.15% and 94.13%, and F1-scores of 99.50% and 96.89%, respectively (mean over three independent training runs). For the subset of correctly matched detections (true positives, determined via a distance-based matching criterion with a 5-pixel tolerance), the localization error was pixel-level, with a mean below 1 px, effectively converting coarse bounding boxes into precise pixel coordinates. Furthermore, our analysis revealed that the deep learning model demonstrates context-aware behavior by intelligently disregarding imaging artifacts, such as vessel truncations at image borders, which traditional morphological methods often misclassify as anatomical endpoint(s). This study presents a promising framework for extracting topological features from vascular networks, combining the contextual awareness of deep learning with the precision of mathematical morphology, though broader validation across datasets, species, and imaging modalities remains necessary to establish generalizability.