What is it about?
Thyroid cancer comes in different forms, with papillary thyroid carcinoma (PTC) and follicular thyroid carcinoma (FTC) being the most common. These two types require different treatment approaches: FTC often needs complete removal of the thyroid, while some low-risk PTC cases can be managed less aggressively. However, telling them apart under a microscope can be challenging even for experienced pathologists, because their capsules – the collagen-rich outer layers surrounding the nodules – can look similar or vary greatly within the same tumor. In this study, we used a special label-free imaging technique, second harmonic generation (SHG) microscopy, to visualize these collagen capsules in detail, without the need for dyes or stains. We then extracted hundreds of texture and intensity features from the images and trained several machine learning models to automatically distinguish between PTC and FTC. One of the biggest challenges was "label noise" comprised of areas like calcifications or surrounding normal tissue that were mislabeled or didn't reflect the tumor type. By using unsupervised clustering to separate these misleading regions and carefully selecting the most relevant features, we achieved an accuracy of nearly 85% with a support vector machine classifier. This shows that combining SHG imaging with machine learning could become a reliable tool to help pathologists make more accurate diagnoses and improve patient care.
Featured Image
Photo by Joshua Sortino on Unsplash
Why is it important?
This research is important because it addresses a real clinical challenge: the misclassification of PTC and FTC can lead to overtreatment of low-risk patients or undertreatment of more aggressive cancers, affecting patient outcomes and quality of life. What makes our approach unique is the combination of wide-field SHG microscopy, which allows rapid imaging of entire tissue sections rather than small, potentially unrepresentative regions, with advanced machine learning and careful noise handling. While other studies have looked at small regions of interest, our method captures the full heterogeneity of the tumor capsule, which is a critical diagnostic feature. We also systematically addressed label noise, which is a common but often overlooked problem in medical imaging, by using unsupervised segmentation to separate normal tissue and calcifications from the capsule itself. Our findings demonstrate that with proper data preprocessing and model selection, SHG-based machine learning can achieve clinically relevant accuracy (84.73%), significantly outperforming standard approaches. This work lays the foundation for a practical, automated adjunct to conventional histopathology, potentially reducing diagnostic errors and supporting personalized treatment decisions.
Perspectives
This project was driven by a pressing clinical problem: the need for better tools to distinguish thyroid cancer types and avoid unnecessary surgeries. What I found particularly interesting was how label noise (misleading data points from calcifications or surrounding tissue) could drastically affect model performance, and how addressing it through unsupervised segmentation made such a difference. It was a valuable lesson that data quality often matters as much as the algorithm itself. Another highlight was seeing how the support vector machine (C-SVC) consistently outperformed more complex ensemble models, reminding us that simpler approaches can be more robust when data are noisy and limited. Collaborating with pathologists and imaging experts across Lithuania and Romania also reinforced the importance of interdisciplinary work in translating research into clinical practice. Overall, this study not only demonstrates a promising diagnostic tool but also provides practical insights into handling real-world data challenges in medical imaging, which I hope will guide future work in this area.
Dr. Lena N Golubewa
State research institute Center for Physical Sciences and Technology (Valstybinis mokslinių tyrimų institutas Fizinių ir technologijos mokslų centras (FTMC))
Read the Original
This page is a summary of: Supervised Machine Learning Thyroid Carcinoma Diagnosis Using Wide-Field SHG Microscopy, IEEE Access, January 2025, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/access.2025.3583435.
You can read the full text:
Contributors
The following have contributed to this page
