Visible Light-Near Infrared Hyperspectral Imaging and Deep Learning Enable Rapid, Non-Staining Assessment of Lung Adenocarcinoma

Abstract

Accurate identification of driver mutations such as ALK, EGFR, and KRAS in lung adenocarcinoma is essential for guiding personalized therapies, yet standard genomic assays are invasive and may alter tissue integrity. In this study, we introduce a non-destructive genotyping approach that combines visible-to-near–infrared hyperspectral imaging (400–1000 nm) of unstained pathological sections with a dual-branch deep-learning fusion framework and gradient-boosting classification. The imaging system captures rich spectral–spatial signatures, which are processed by a fusion network that synergistically extracts global contextual features and local textural details. These fused representations are then classified by an optimized XGBoost model. Evaluation on 90 clinical specimens yielded class-specific accuracies between 83.5% and 90.2%, and area under the ROC curve values from 0.83 to 0.91. Our results demonstrate that hyperspectral imaging coupled with deep-learning fusion enables rapid, tumor genotyping, offering a promising tool for real-time clinical diagnostics in the field of biomedical photonics.