Poster #63 - Irina Zhu

AI-driven 3D Digital Pathology of Immune Follicular Structures in Human

Zhu, Irina Y., MS (Department of Bioengineering, University of Pennsylvania; Oldridge Lab, Children's Hospital of Philadelphia, Philadelphia, PA, USA) Dubensky, Sam B., BS (Immunology Graduate Group, University of Pennsylvania; Oldridge Lab, Children's Hospital of Philadelphia, Philadelphia, PA, USA) Carter, Ashley, MS (Oldridge Lab, Children's Hospital of Philadelphia, Philadelphia, PA, USA) Cabrera, Emylette C., BS (Romberg Lab, Children's Hospital of Philadelphia, Philadelphia, PA, USA) Baxter, Amy E., PhD (Department of Systems Pharmacology and Translational Therapeutics, Immune Health, University of Pennsylvania, Philadelphia, PA, USA) Romberg, Neil D., MD (Division of Allergy and Immunology, Children's Hospital of Philadelphia; Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA) Oldridge, Derek A., MD, PhD (Department of Pathology and Laboratory Medicine, Children's Hospital of Philadelphia; Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA)

Reactive germinal centers (GCs) in secondary lymphoid organs orchestrate humoral immunity. However, their complex three-dimensional (3D) architecture is destroyed by bulk or dissociative single-cell methods and is often missed or distorted by 2D spatial methods. We present an end-to-end, AI-driven pipeline that reconstructs 3D lymphoid microanatomy from serial H&E whole-slide images (WSIs). In a pilot of 40 consecutive tonsil sections spanning approximately half a millimeter (~0.42mm), we combined robust WSI registration with pathology AI-guided tissue segmentation to reconstruct comprehensive 3D digital tonsil volumes. Using these AI-driven approaches, we were able to delineate key tissue microanatomy, including follicles, GCs, mantle, inter-follicular, epithelial, and vascular compartments with minimal manual annotation. This automated segmentation approach achieved 90.1% accuracy against pathologist annotations, capturing major follicular structures and supporting downstream 3D analysis in a highly cost-effective and scalable manner. Upon reconstruction of the 3D tonsil, we identified 117 unique follicles. Combining qualitative visualization with quantitative morphometrics across all follicles, we found that follicles were predominantly ellipsoidal; however, their volumes varied considerably across the tonsil (mean ~0.141 mm³; range 0.014 - 0.467mm³). Notably, several follicles that appeared to lack a GC on single 2D sections contained GC tissue elsewhere in the 3D volume, underscoring how 2D sampling can misclassify follicular age. Weaving together computational methods from spatialomic neighborhood analysis and pathology AI, we were further able to resolve fine cellular detail within the GC (including centroblasts, centrocytes, and tingible-body macrophages) as well as contiguous 3D light- and dark-zone domains consistent with expected GC polarization. This reproducible workflow enables high-throughput 3D visualization and quantification of the lymphoid structures. Our future work will extend this pipeline to lymph nodes in diseases with follicular pathology, including common variable immunodeficiency and follicular lymphoma. We will also integrate spatial transcriptomics along with morphology to interrogate 3D cell-cell communication underlying immune dysregulation.