Thyroid cancer is the most common endocrine cancer, affecting more people each year as detection rates continue to rise. During tumor excision, surgeons often struggle to determine exactly how much tissue should be removed, as distinguishing cancer from healthy tissue in real time is challenging and nearby structures are extremely delicate. Today, diagnosis and margin assessment rely on fine-needle aspiration (FNA) and traditional pathology. While accurate, these methods are slow, sometimes inconclusive, and offer no real-time guidance in the operating room. As a result, patients may undergo unnecessary surgery for benign nodules or require additional procedures if cancerous tissue is missed.
Dynamic Optical Contrast Imaging (DOCI) offers a fundamentally different way to examine tissue. Rather than using dyes or contrast agents, DOCI illuminates tissue and measures its natural autofluorescence, the faint light emitted by molecules already present in cells. Healthy and cancerous tissue glow differently, creating optical signatures that can be captured and analyzed. Each DOCI scan collects information from 23 different optical channels across a wide field of view, producing a rich spectral fingerprint detailed map of tissue biology in freshly excised specimens.
In a multi-institutional study published in Biophotonics Discovery , researchers combined DOCI with machine learning to translate these complex optical signals into clinically meaningful insights. At Duke University, Tyler Vasse, working in the laboratory of Tuan Vo-Dinh, developed a two-stage AI analysis framework. Clinical deployment and application of the DOCI technology was carried out by Yazeed Alhiyari and team members of Maie St. John at the University of California, Los Angeles.
In the first stage, the researchers employed a simple, interpretable machine-learning model to assign each specimen to one of three categories: healthy thyroid tissue, follicular thyroid cancer, or papillary thyroid cancer, the latter two representing the most common differentiated thyroid malignancies. By distilling the 23 DOCI optical channels into a small set of key features, the system accurately classified samples across these categories, achieving perfect accuracy on an independent test set. Notably, the model also correctly identified samples from the highly aggressive anaplastic subtype as cancerous, demonstrating broad sensitivity to malignant tissues.
The second stage addressed a critical surgical question: where exactly is the tumor located? To answer this, the team employed deep-learning models based on a U-Net architecture, which is designed to identify and map specific regions within medical images. Using this approach, the models generated tumor probability maps that accurately highlighted cancerous regions, with particularly strong performance for papillary thyroid cancer, and very low false-positive rates in cancer-free tissue.
Although this work analyzed tissue immediately after removal, the results point toward a future in which surgeons could receive rapid, label-free guidance during operations. By merging the speed of optical imaging with the power of AI, DOCI has the potential to reduce uncertainty in the operating room, prevent unnecessary surgeries, spare healthy tissue, and improve outcomes for patients with thyroid cancer, allowing surgeons to see cancer in—quite literally—a new light.
For details, see the original Gold Open Access article by T. Vasse et al., " Machine-learning-based tumor segmentation and classification using dynamic optical contrast imaging for thyroid cancer ," Biophoton. Discovery 3(1), 105001 (2026), doi: 10.1117/1.BIOS.3.1.015001
