Comparative Analysis of Deep Learning and Traditional Clustering Methods for Thermal Thyroid Image Segmentation: A Clinical Evaluation Study

Abstract

Thermal imaging has emerged as a revolutionary non-invasive diagnostic modality for thyroid disorders, offering clinicians the ability to visualize real- time temperature distributions without exposing patients to harmful radiation. The accuracy of thyroid region segmentation in thermal images remains a critical bottleneck for developing reliable automated diagnostic systems and enhancing clinical assessment capabilities. This comprehensive study presents a rigorous comparative analysis of three distinct segmentation methodologies for thermal thyroid imaging: a simplified U-Net deep learning architecture, K-Means clustering with adaptive parameters, and Fuzzy C-Means (FCM) clustering with enhanced robustness features. Our evaluation framework incorporates multiple performance metrics alongside clinically relevant parameters to provide a holistic assessment of each approach. This paper specifically highlights areas where MATLAB-generated results and code can be integrated to enhance the reproducibility and clarity of the presented methodologies and findings.