FigATree: a novel framework for histological subtyping and grading of lung adenocarcinoma

Lung adenocarcinoma (LUAD) exhibits pronounced morphological heterogeneity, making accurate subtyping and grading a persistent challenge in pathology. Conventional deep learning methods often lack the granularity and interpretability required for clinical translation. We introduce FigATree, an interpretable AI framework for LUAD diagnosis that combines a foundation model-enhanced region-level encoder with an XGBoost-based pathology-informed slide-level classifier. Applied to 1186 H&E-stained whole-slide images, FigATree achieves $>95\%$ accuracy in classifying six histological patterns at the region level and  90% and  85% accuracy in slide-level subtyping and differentiation grading, respectively, substantially outperforming guidelines-based baselines. Additionally, FigATree achieved region-level and slide-level accuracies of nearly 100% and 80%, respectively, on external validation, fully demonstrating its generalizability. The framework yields interpretable predictions aligned with pathological criteria, offering transparency at both region and slide levels. By integrating foundation model representations with a clinically grounded decision module, FigATree enables accurate, explainable classification of LUAD and represents a scalable paradigm for computational histopathology.