Deep learning framework for real-time label-free thyroid cancer detection using second harmonic generation microscopy

Fine needle aspiration (FNA) is the primary diagnostic tool for thyroid cancer but has key limitations, including a 21% false-negative rate and 30% indeterminate results, often leading to repeat biopsies, expensive molecular testing, and unnecessary surgeries. Second harmonic generation (SHG) microscopy, a nonlinear optical technique, is being studied as a label-free method of diagnosing thyroid cancer due to its ability to detect malignancy-associated changes in collagen architecture. Although computational methods have demonstrated potential in distinguishing normal from cancerous thyroid tissue, integrating SHG with artificial intelligence (AI)-based analysis can greatly improve diagnostic speed and accuracy. In this study, SHG images acquired from normal and cancerous thyroid sections were analyzed using various deep learning models to assess their diagnostic performance and identify the optimal architecture for clinical deployment. Results showed that DenseNet-121 achieved superior performance with 96.3% AUC (81.3% sensitivity, 92.8% specificity), while ensemble methods reached 96.6% AUC, with all models demonstrating real-time inference capabilities (12ms). These findings suggest that AI-enabled SHG represents a clinically viable alternative to invasive biopsies, offering immediate, accurate thyroid cancer detection at the point of care.