Machine Learning-Based Tumor Segmentation and Classification Using Dynamic Optical Contrast Imaging (DOCI) for Thyroid Cancer

Abstract

Thyroid cancer presents significant diagnostic challenges due to its complex anatomy and diverse tissue types. This study leverages Dynamic Optical Contrast Imaging (DOCI), a label-free, real-time imaging technology, with machine learning to enhance tumor detection and segmentation. Using 23 DOCI filters, we applied Principal Component Analysis (PCA) for dimensionality reduction, k-nearest neighbors (k-NN) for classification, and U-Net models for segmentation. The approach achieved high accuracy in distinguishing tissue types, with PCA enabling clear clustering, k-NN classifying normal, follicular, and papillary tissues, and U-Net models achieving 96.34% and 92.02% accuracy for papillary and follicular segmentation, respectively. Filter importance analysis reduced input dimensionality without significantly compromising performance, highlighting the potential for optimized imaging protocols. These findings demonstrate DOCIs utility in improving diagnostic accuracy and tumor characterization in thyroid cancer and beyond, offering a foundation for personalized treatment planning and surgical precision.