Label-Free Microscopic Hyperspectral Imaging for Papillary Thyroid Carcinoma Photodiagnosis Using Patient-Level Spectral-Spatial Deep Learning.

BACKGROUND

Papillary thyroid carcinoma (PTC) may overlap microscopically with Hashimoto thyroiditis (HT) and nodular goiter (NG). Label-free optical assessment that captures morphology and wavelength-dependent tissue contrast could provide an objective adjunct to conventional pathology.

METHODS

We retrospectively analyzed 180 patients (72 PTC, 54 HT, and 54 NG). Unstained serial sections were imaged by visible-near-infrared microscopic hyperspectral imaging (400-1000 nm; effective range 420-950 nm), with adjacent hematoxylin-eosin sections as the reference. We developed Thyroid Hyperspectral Imaging Multiple-Instance Learning (ThyroHSI-MIL), a patient-level dual-branch framework combining multi-scale spectral encoding, wavelength attention, spatial feature extraction, gated spectral-spatial fusion, and attention-based multiple-instance aggregation. After white-dark calibration, Savitzky-Golay smoothing, and standard normal variate normalization, model performance was assessed by stratified five-fold patient-level cross-validation.

RESULTS

For PTC versus benign lesions, ThyroHSI-MIL achieved an AUC of 0.927, accuracy of 0.856, sensitivity of 0.875, specificity of 0.843, and F1 score of 0.842. Three-class classification achieved a macro-average AUC of 0.903 and accuracy of 0.844. The strongest wavelength-attention peaks occurred near 468 and 530 nm. Ablation analysis supported contributions from the spectral and spatial branches, wavelength attention, gated fusion, and patient-level aggregation.

CONCLUSIONS

Label-free microscopic hyperspectral imaging with patient-level spectral-spatial learning accurately differentiated PTC from benign thyroid lesions and provided wavelength- and region-level interpretability. External and prospective validation is warranted before clinical deployment.