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

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Pathology

Background:

• Thyroid nodules are common, but differentiating malignant from benign types is challenging with current methods.
• Traditional diagnosis involves invasive biopsies and lengthy histopathology, hindering real-time assessment.
• Optical coherence tomography (OCT) offers non-invasive imaging, but its pathological interpretation is difficult.

Purpose of the Study:

• To develop a deep learning (DL) framework for classifying thyroid nodule pathology using OCT images.
• To enable accurate, real-time pathological assessment of thyroid nodules non-invasively.
• To improve diagnostic accuracy for thyroid carcinoma subtypes and benign conditions.

Main Methods:

• Acquired OCT datasets from seven pathological categories (5 carcinoma subtypes, benign, normal).
• Utilized histology-matched OCT data for supervised deep learning model training.
• Developed a DL framework for binary (carcinoma vs. non-carcinoma) and multi-class classification.
• Visualized diagnostic predictions using color-coded overlays on OCT images.

Main Results:

• Achieved 98.37% accuracy and 0.997 AUC for binary classification of carcinoma vs. non-carcinoma.
• Demonstrated 93.66% overall accuracy for multi-class classification across seven categories.
• Enabled coherent interpretation of tissue pathology through visualized diagnostic predictions.

Conclusions:

• The combination of OCT and DL shows feasibility for enhanced thyroid pathology assessment.
• This approach supports real-time and point-of-care diagnostic applications for thyroid nodules.
• Further optimization can lead to improved clinical deployment for thyroid nodule diagnostics.