Label-free screening and grading of follicular thyroid neoplasms enabled by Fourier transform infrared microspectroscopy and machine learning.

Follicular thyroid neoplasms represent a common subtype of thyroid tumor that has become the most prevalent endocrine neoplasms in recent decades. Accurate diagnosis and grading of these tumors are critical for clinical management of follicular thyroid neoplasms, yet remain challenging due to the limitations of conventional imaging modalities without histopathological molecular information. There is growing interest in analytical techniques that can provide metabolic and molecular insights without the need for exogenous reagents. In such a context, Fourier transform infrared (FTIR) microspectroscopy has emerged as a promising label-free approach for detecting intrinsic disease biomarkers. In this work, we proposed a label-free framework for screening and grading follicular thyroid neoplasm tissues using FTIR microspectroscopy and machine learning. A total of 32 clinical samples in the form of tissue sections were collected from a real-world cohort and measured for the spectral data, from patients diagnosed with follicular thyroid adenoma, follicular tumor with uncertain malignant potential, and follicular thyroid carcinoma. For tumor screening, disease-specific features were extracted from the FTIR mapping data and further imaged through integration, principal component analysis, and clustering, enabling visual and quantitative assessment of tissue abnormalities. For neoplasms grading, a deep neural network trained with an adversarial learning strategy achieved a grading accuracy of 94.0% on an independent test set. These findings collectively demonstrate the potential of FTIR microspectroscopy as a powerful, reagent-free tool for the diagnosis, pathological evaluation, and clinical management of follicular thyroid neoplasms.