Papillary thyroid carcinoma (PTC) and nodular goiter with papillary hyperplasia (NGPH) share similar histological features, complicating both preoperative and intraoperative diagnoses. We assessed hyperspectral imaging (HSI) combined with deep learning to differentiate PTC from NGPH. Forty-three paraffin-embedded PTC and 39 NGPH samples were imaged across 400-1000 nm, with reflectance calibration and Savitzky-Golay smoothing applied. Extracted spectral features were input into a one-dimensional convolutional neural network with a self-attention mechanism. HSI demonstrated sensitivity above 90% in the 500-600 nm and near-infrared regions for distinguishing PTC and NGPH. The model achieved an area under the ROC curve of 0.8635 and pixel-level classification accuracy of 87.07%, with both sensitivity and specificity at 87%. Spectral feature depth correlated significantly with histopathological parameters. These findings indicate that HSI combined with deep learning can accurately capture spectral differences between PTC and NGPH, supporting its potential for rapid intraoperative guidance and noninvasive preoperative screening.
Keywords: deep learning; hyperspectral imaging; papillary thyroid cancer; spectral analysis.
© 2025 Wiley‐VCH GmbH.