Image: AI models combined with DOCI can classify thyroid cancer subtypes (photo courtesy of T. Vasse et al., doi 10.1117/1.BIOS.3.1.015001)
Thyroid cancer is the most common endocrine cancer, and its increasing detection rate has increased the number of patients undergoing surgery. During tumor removal, surgeons are often faced with the uncertainty of distinguishing cancerous tissue from healthy structures, increasing the risk of incomplete excision or unnecessary surgery. Current diagnostic tools, such as fine needle aspiration and postoperative pathology, are accurate but slow and do not offer real-time guidance in the operating room. New research demonstrates a label-free optical imaging approach, combined with artificial intelligence (AI), that can quickly identify and localize thyroid cancerous tissue.
Researchers at Duke University (Durham, North Carolina, USA) and the University of California, Los Angeles (UCLA, Los Angeles, CA, USA) used Dynamic Optical Contrast Imaging (DOCI), a technique that illuminates tissue and measures its natural autofluorescence instead of using dyes or contrast agents. Each DOCI scan captures data from 23 optical channels over a wide field of view, generating a detailed spectral fingerprint that reflects the underlying tissue biology.
Freshly resected thyroid samples were imaged using DOCI and analyzed with a two-stage machine learning system. In the first stage, an interpretable classification model distilled complex optical data into a small set of features to categorize tissue as healthy, follicular thyroid cancer, or papillary thyroid cancer. In the second stage, deep learning models based on the U-Net architecture were applied to generate spatial maps of tumor probability, identifying the precise location of cancerous regions in each sample.
The findings, published in Biophotonics Discoveryshow that the AI system accurately classified thyroid tissue types and achieved perfect accuracy in an independent test set. Notably, it also correctly identified aggressive anaplastic thyroid cancer samples as malignant, despite not having been explicitly trained for that subtype. Deep learning segmentation models generated highly accurate tumor maps, particularly for papillary thyroid cancer, while maintaining very low false positive rates in non-cancer tissue.
Although the current study analyzed tissue after surgical removal, the results point to its intraoperative use in the future. By providing rapid, label-free visualization of cancer margins, DOCI combined with AI could help surgeons remove tumors more precisely, reduce re-interventions, and preserve healthy tissue. With further development, this approach could offer real-time guidance in the operating room, thereby improving outcomes for patients with thyroid cancer.
Related links:
Duke University
University of California, Los Angeles (UCLA)
