Bio-interpretable ensemble learning model for invasive pulmonary adenocarcinoma grade using CT and histopathology images

Abstract The significant heterogeneity and complex morphology of invasive pulmonary adenocarcinoma (IPA) make grading challenging for pathologists. However, thorough investigations into radiopathomics features extracted from computed tomography (CT) and whole slide images (WSIs) for IPA grading and their biological significance remain limited. We aim to integrate multi-omics analysis to establish a robust grading model for IPA and reveal its biological significance. This multicenter study encompassed 988 patients who underwent radical surgical resection and received a pathological confirmation of IPA. Through integrated analysis of radiomics and pathomics, we constructed and validated an optimal ensemble learning grading model, which integrates multi-scale and multi-modal characteristics, achieved AUCs of 0.885, 0.920, 0.833, and 0.905 in the internal and external validation sets. Further systematic analysis of paired CT, WSIs, and RNA sequencing, two co-expression modules, 23 hub genes, and 680 significant pathways associated with grading were identified. Moreover, the reproducibility of the radiopathomics phenotypes, linked to multiple biological pathways—including signal transduction, cell differentiation, DNA damage and repair, cell proliferation and growth, metabolism, and metastasis and invasion—has been validated. In conclusion, the integration of radiological and pathological characteristics enhances the accuracy in differentiating high-grade IPA, offering a robust approach for grading. Multi-scale imaging biomarkers may promote personalized treatment.