Applications of artificial intelligence in non-small cell lung cancer: from precision diagnosis to personalized prognosis and therapy.

[BACKGROUND] Non-small cell lung cancer (NSCLC) carries a major global burden. The rapid growth of multimodal medical data challenges conventional methods to deliver stable, transferable and interpretable decisions across heterogeneous longitudinal high dimensional inputs.

[METHODS] This review summarizes advances in artificial intelligence (AI) for NSCLC from 2023 to 2025 and outlines a translation-focused framework that links algorithmic progress to clinical utility. We survey thoracic imaging, digital pathology and multiomics together with evaluation practices and implementation guidance. We also adopt a critical perspective.

[RESULTS] Many high performing deep models remain black boxes, and popular post hoc explanations such as Grad CAM heatmaps are rarely validated for faithfulness or stability, which undermines clinician trust and limits use in high stakes decisions. To address this gap, we propose a minimum evidence package for explainability that comprises sanity checks, quantitative faithfulness tests such as deletion or insertion, ROAR or IROF and infidelity, stability analyses, concept level validation for example TCAV with statistical testing, and prospective human factors studies that demonstrate improved decisions without automation bias. Across modalities, evaluation has expanded beyond discrimination to include calibration, uncertainty quantification (UQ) and subgroup analyses across scanners, sites and populations. However, the evidence base remains constrained by retrospective single center designs, inconsistent external or temporal validation and limited decision curve analysis (DCA). Translational priorities include a staged validation ladder from technical to clinical to prospective deployment, alignment with Software as a Medical Device frameworks, interoperable governance, fairness and economic assessment, and validated explainability coupled with uncertainty aware selective workflows.

[CONCLUSIONS] Looking ahead, progress will depend on multimodal foundation models, causal and temporal modeling, and regulatory qualification of computable biomarkers with verified explanations, supported by multicenter prospective studies that demonstrate durable generalizability, clinical value and clinician trust.