Multimodal multitask deep learning for grading management system in non-small cell lung cancer

Accurate histologic subtyping, tumor node metastasis classification (TNM) staging and prognostic assessment are central to clinical management of non-small cell lung cancer (NSCLC), but remain challenging because of tumor heterogeneity, limited biomarker performance and diagnostic uncertainty. Here we show that a multimodal, multi-task deep learning scoring system (MM-DLS), integrating pretreatment PET/CT images with clinical variables, enables non-invasive prediction of NSCLC subtype, stage and survival risk. We develop and validate MM-DLS in 4,164 patients from multiple centres. In the external validation cohort, MM-DLS achieves area under the receiver operating characteristic curve values of 0.86 for histologic subtype classification and 0.86, 0.86, and 0.88 for stages I-II, III, and IV, respectively. The model also shows consistently strong discrimination for 1-, 3-, and 5-year survival across treatment regimens. These results indicate that MM-DLS provides an integrated framework for subtype prediction, staging, and prognostic stratification in NSCLC.