Label-free diagnosis across the thyroid nodule pathology spectrum using deep learning-enabled optical coherence tomography

Thyroid nodules are highly prevalent, yet identifying malignancy remains a persistent challenge due to their significant pathological heterogeneity. Conventional diagnostic workflows rely on invasive and tissue-destructive sampling followed by a time-consuming histopathology tissue process, restricting the ability to obtain pathological insights in real-time or at the bedside. In this context, optical coherence tomography (OCT) has gained attention as a non-invasive, label-free imaging modality; however, its interpretation for detailed pathological assessment has remained challenging. Here, we developed a deep learning (DL)-based framework for diagnostic classification across the spectrum of thyroid nodule pathology using OCT images. OCT datasets were acquired from seven pathological categories, including five thyroid carcinoma subtypes as well as two non-carcinoma tissue types (benign and normal), and were matched with histology for supervised learning. Robust binary differentiation between carcinoma and non-carcinoma was achieved, with an accuracy of 98.37% and an area under the receiver operating characteristic curve of 0.997. Furthermore, multi-class classification across seven pathological categories further demonstrated an overall accuracy of 93.66% on held-out test sets. Diagnostic predictions were visualized as color-coded overlays on e n face OCT images, enabling coherent interpretation across tissue volumes. These results demonstrate the feasibility of combining OCT with DL for enhanced thyroid pathology assessment and motivate further optimization toward real-time and point-of-care diagnostic deployment.