Grading of Castleman Disease Histopathology with an Attention-Based Multiple Instance Learning Model

Background/Objectives: Castleman disease is a rare cytokine-driven lymphoproliferative disorder in which lymph node histopathology provides key diagnostic information. Morphologic features are graded semiquantitatively and often show substantial interobserver variability. Methods: We developed an automated approach to grade six Castleman disease-associated histologic features on hematoxylin and eosin (H&E) whole slide images (WSIs) using an attention-based multiple instance learning (MIL) model built on a large pathology foundation model encoder. A multi-institution cohort comprised 544 lymph node WSIs, including 397 from cases with Castleman disease or Castleman-like histology and 147 from cases without suspicion for Castleman disease. These case ascertainment categories were not model prediction targets. Slides were assigned ordinal grades (0–3) for regressed germinal centers, follicular dendritic cell prominence, increased vascularity, hyperplastic germinal centers, plasmacytosis, and follicular twinning by hematopathologists. Model performance was assessed on a held-out evaluation set of 142 WSIs. Results: Across the six features, accuracy ranged from 0.52 to 0.68 (mean 0.60). Disagreements were predominantly minor: 96% of predictions were within one grade of the reference. Feature-specific tile contribution heatmaps showed qualitative spatial correspondence with selected plasma cell-rich, vessel-rich, and follicular regions but were not evaluated as quantitative feature localization maps or causal explanations. In a preliminary reader study, concordance with the reference varied widely among hematopathologists (Krippendorff’s alpha 0.20–0.95, mean 0.50), with the model demonstrating concordance in the mid-range (0.56). Conclusions: These findings support the feasibility of automated grading of Castleman disease-associated histologic features for research standardization. The model does not classify Castleman disease, and its potential use as an input to diagnostic or clinical trial workflows requires evaluation in separately designed studies with adjudicated diagnostic labels and integrated clinical data.