Poster #29: Alicia Holk MS

Rapid Segmentation of High-Resolution Microscopy Images Using an Integrated Machine Learning Approach

Alicia Holk .M.S.1, Margaux Deputy2, Michael Donzanti2 , Logan Hallee, Ph.D.1, Jason P. Gleghorn, Ph.D.2 1Center for Bioinformatics & Computational Biology, University of Delaware, Newark, DE, USA 2Department of Biomedical Engineering, University of Delaware, Newark, DE, USA

Lymph nodes can serve as sanctuary sites for diseases such as cancer and HIV where vascular barriers limit therapeutic delivery. Volumetric quantification of vessels, follicles, and lobules within the lymph node have played an essential role in improving predictions of drug delivery rates and dosages; however, limited annotated datasets are available. Manual image annotation can require 4-6 hours per image, limiting sample throughput and delaying downstream analyses. Although existing segmentation tools such as FIJI and Segment Anything Model 3 (SAM3) improve the accessibility of image annotation, these approaches still struggle to segment complex vasculature and larger anatomical structures in a timely and accurate manner. Specialized machine learning (ML) models can segment these complexities with relatively high fidelity, but require large datasets of pre-segmented masks. To address this bottleneck, we developed a two-stage integrated workflow that tackles both the ground-truth annotation barrier and the need for scalable, high-throughput segmentation. To reduce the time required to prepare an ML model training set, we incorporated FIJI for vessel segmentation, SAM3 via Roboflow for large anatomical structure segmentation, and Adobe Photoshop for final quality control and correction. Overall, the method reduced initial image annotation to an average of 20 minutes per image. The images and their resulting ground truth were then used to train a specialized multi-label ML model capable of segmenting an entire dataset of 200 high-resolution microscopy images in less than 5 minutes. The final ML model demonstrated robust validation metrics including a mean Matthews Correlation Coefficient (MCC) of 0.865 and a mean F1 score of 0.870. With the ability to segment hundreds of high-resolution lymph node images in minutes, we can now produce expansive datasets that can be used for drug delivery analyses across a variety of anatomical and pathological states.