Standardized Instance-Level Quantification of CD34-Positive Vessels in Lymph Node Whole-Slide Images Using U-Net.

Deep learning enhances quantitative analysis of blood vessels in whole-slide images (WSIs), leading to precise and reliable counting. However, counting vessels in lymph nodes-which contain distinctive high endothelial venules (HEVs)-has not been systematically evaluated. Artificial intelligence (AI) approaches to vessel-counting fall into two categories: semantic segmentation, which provides a continuous mask, and instance segmentation, which treats each vessel as a separate object; prior work has emphasized the former. To address this gap, we adopted an instance-level approach to detect and count individual vessels. We developed a convolutional neural network (CNN) to quantify CD34-positive vessels in lymph-node WSIs, using U-Net, a widely used biomedical segmentation architecture. The model was trained over 4.0 million (4.0 M) iterations, followed by a fixed post-processing procedure to extract vessel instances. At 3.0 M, the model most faithfully captured CD34-positive vascular morphology. In the validation set, this checkpoint showed the highest morphological similarity. Furthermore, detection performance was robust with minimal variation across lymph nodes and training checkpoints. Using a simple fixed procedure, our method achieves accurate vessel detection and counting in lymph-node WSIs. This model provides a practical basis for standardized quantification and for studying lymph-node vascular morphology across normal, inflammatory, and lymphomatous states.