Introduction
Microcirculation—the network of capillaries, arterioles, and venules—is essential for organ health and serves as a vital indicator of disease progression. In clinical research, Intravital Video Microscopy (IVM) is the standard tool for observing these vessels. However, the analysis of IVM data remains a significant bottleneck. Experts must manually identify hundreds of vessels and determine their functionality (whether blood is flowing) through tedious visual inspection. This manual process is not only slow and expensive but also subjective, often requiring multiple independent reviewers to ensure consistency. This paper presents an automated two-step machine learning framework designed to perform simultaneous structural and functional microcirculation analysis with high accuracy.
Methodology
The proposed pipeline automates the transition from raw video data to a functional vessel map through two distinct stages: segmentation and flow determination.
Vessel Segmentation
The first stage identifies the vessel network structure. To highlight active areas and filter out static tissue, the system generates a Sum of Absolute Differences (SAD) image, which collapses the motion within a video sequence into a single 2D frame. A modified Steger Unbiased Detector—a classical curvilinear feature extractor—is then applied to this motion-summarized image. Combined with morphological filters, this approach maps the spatial coordinates of the vessel network.
Functional Flow Analysis
Once the vessels are located, the system determines if they are functional. Because capillary flow often involves discrete red blood cells separated by plasma gaps, a 2D snapshot is insufficient for determining movement. The researchers utilized a 3D Convolutional Neural Network (3D-CNN) to incorporate the temporal dimension.
For every pixel in a detected vessel's cross-section, the algorithm extracts a 16x16x200 data block representing a spatial area over time. This 3D block is processed by the CNN, which uses two hidden convolution layers (7x7x7 and 5x5x5) followed by max-pooling and a Multi-layer Perceptron (MLP) to classify the pixel as "flowing" or "not flowing." A vessel is deemed functional if more than 50% of its cross-sectional pixels are flagged as flowing.
Experimental Results
The framework was evaluated using a dataset of 458 IVM videos of rat muscle tissue, comprising over 32,000 labeled vessels. To ensure robust evaluation, the data was split by video rather than by individual vessel to prevent the model from memorizing specific video backgrounds.
- Classification Accuracy: The 3D-CNN achieved 90% accuracy on the held-out test set.
- Reliability: The model demonstrated high precision (0.93) and sensitivity (0.95), indicating it rarely misses functional vessels or misidentifies flow.
- End-to-End Performance: In real-world deployment on unlabeled videos, the system captured 83% of all vessels identified by human experts, with a remarkably low false positive rate of 3.5%.
Discussion and Future Work
The hybrid approach—combining classical computer vision for segmentation with deep learning for temporal analysis—proves highly effective for microcirculation. The 3D-CNN successfully filters out non-vessel artifacts that might be erroneously picked up during the initial segmentation phase.
The primary limitation identified was the classical Steger detector, which occasionally struggled with intersecting vessels or blurry outlines in deep tissue. Future iterations of this work aim to replace the classical segmentation step with a dedicated deep-learning-based segmentation model to further improve the system's vessel capture rate and overall robustness.
Summary updated