Artificial Intelligence-assisted thyroid cancer diagnosis from Raman spectra of histological samples

Raman spectroscopy emerges as a highly promising diagnostic tool for thyroid cancer due to its capacity to discern biochemical alterations during cancer progression. This non-invasive and label/dye-free technique exhibits superior efficacy in discriminating malignant features compared to traditional molecular tests, thereby minimizing unnecessary surgeries. Nevertheless, a key challenge in adopting Raman spectroscopy lies in identifying significant patterns and peaks. This study proposes an artificial intelligence approach for distinguishing healthy/benign from malignant nodules, ensuring interpretable outcomes. Raman spectra from histological samples are collected, and a set of peaks is selected using a data-driven, label-independent approach. Machine Learning algorithms are trained based on the relative prominence of these peaks, achieving performance metrics with an area under the receiver operating characteristic curve exceeding 0.9. To enhance interpretability, eXplainable Artificial Intelligence (XAI) is employed to compute each feature's contribution to sample prediction.