Application of Hyperspectral Imaging and Machine Learning for Differential Diagnosis of Hashimoto's Thyroiditis and Papillary Thyroid Carcinoma

Abstract

Background

Hashimoto's thyroiditis (HT) and papillary thyroid carcinoma (PTC) often share similar features, leading to frequent misdiagnoses. Hyperspectral imaging (HSI) offers detailed spatial and spectral insights, promising improved tumor detection.

Objective

This study aims to discern HT and PTC spectral characteristics using HSI and evaluate deep learning models for pathologic diagnostic effects.

Methods

Hyperspectral data from HT and PTC samples were processed using second-order derivatives and Savitzky–Golay smoothing. The adaptive spectral feature selection network model classified spectral data from various wavelengths to assess performance.

Results

PTC showed unique spectral features in the 400–500 nm range with higher peak intensities at lower wavelengths than HT. The model achieved 88.36% accuracy, highlighting the importance of low-wavelength data in differentiating PTC from HT.

Conclusion

The model effectively identifies spectral differences between HT and PTC, offering a novel approach for precise thyroid disease diagnosis.