Intra-operative differentiation of benign and malignant thyroid nodules using diffuse reflectance spectroscopy

Abstract Thyroid nodules are common, yet distinguishing benign from malignant nodules is a challenge. With current diagnostics, 39–52% of nodules are characterised as indeterminate, often necessitating diagnostic surgery to achieve a histological diagnosis. Depending on outcome, this may represent over- or under-treatment. The aim of this study was to determine whether diffuse reflectance spectroscopy (DRS) can accurately differentiate benign from malignant thyroid nodules. In this prospective study, spectral data were collected intra-operatively from freshly resected ex vivo thyroid nodules using a handheld DRS probe; results were validated against histopathological analysis as the gold standard for comparison. Binary classification was performed using the extreme gradient boosting supervised machine learning classifier. Data were divided into training and testing sets with a stratified 5-fold cross-validation method, and model performance evaluated for key statistical features. Our study revealed that for nodules deemed benign or malignant on pre-operative imaging/cytology, DRS identified malignancy with a diagnostic accuracy of 87.9%, sensitivity of 88.3%, specificity of 87.4% and positive predictive value (PPV) for malignancy of 88.0%. In pre-operatively indeterminate nodules, the diagnostic accuracy was 89.0%, sensitivity 79.9%, specificity 93.3% and PPV 85.4%. These findings highlight the potential of DRS in assessing indeterminate thyroid nodules, where its performance exceeds that of current image-guided cytological analysis (where the PPV ranges from 17 to 68%). This study therefore provides justification for further work exploring modification of the DRS system for pre-operative use, in order to support decision-making for indeterminate nodules and reduce unnecessary surgery.