Abstract A021: Leveraging histopathology and clinical information to map the tumor microenvironment for interpretable lung cancer diagnosis

Abstract Lung cancer remains the leading cause of cancer mortality, in part because current screening workflows split clinical findings and surgical pathology, which provides definitive insight but requires invasive biopsy. This disconnect sustains high false-positive rates, overdiagnosis, and delayed follow-ups, underscoring the urgent need for objective, integrative tools that unify histological and clinical information within an interpretable framework. The goal of this work is to develop TMATT, an interpretable machine-learning framework that integrates histopathological images with clinical data to classify lesions and predict unknown pathological information, cancer types, and aggressiveness. Our central hypothesis is that TMATT can capture complementary features across these modalities, yielding robust and clinically meaningful predictions. Building on masked modelling system, TMATT, which achieved 93% AUC for disease classification and a c-index of 0.7 for survival prediction. TMATT implemented this approach by randomly masking both clinical variables and histological image regions and parts of clinical finding during training using transformer architecture. This strategy enables the model to learn latent feature interactions between clinical findings and histology while maintaining interpretability. We have trained and validated TMATT on TCGA and NLST datasets, with generalization extended to multiple organs. The framework outputs lesion classifications (e.g., carcinoma vs. sarcoma vs. lymphoma) and aggressiveness scores (low to high) while generating interpretable heatmaps highlighting the histological regions most relevant to predictions. This work will establish the foundation for multimodal, interpretable, and generalizable lesion classification, serving as a critical step toward noninvasive “virtual pathology” and earlier, more accurate cancer detection. Further evaluation and assessment with pathologists is being conducted to validate the applicabilty of TMATT in clinical setting. Citation Format: Kushal Virupakshappa, Avinash Sahu, David Arredondo, Chris Amos. Leveraging histopathology and clinical information to map the tumor microenvironment for interpretable lung cancer diagnosis [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Cancer Evolution: The Dynamics of Progression and Persistence; 2025 Dec 4-6; Albuquerque, NM. Philadelphia (PA): AACR; Cancer Res 2025;85(23_Suppl):Abstract nr A021.