A Non-Contact Framework for Thyroid Disorder Screening Using Infrared Thermography and GLCM-Based Neural Networks

Thyroid-related disorders are increasingly common and are often identified only after noticeable symptoms appear, which delays timely intervention. This highlights the importance of developing simple and accessible screening approaches. Conventional diagnostic methods, including ultrasound imaging and cytological evaluation is reliable but requires specialized infrastructure and expert interpretation, limiting their use in large-scale or early screening scenarios. This study explores infrared thermography as a non-invasive and radiation-free technique for preliminary thyroid screening. A computationally efficient machine learning framework is developed with a focus on real-time applicability. Initially, thermal images are processed to localize the thyroid region using a U-Net-based segmentation approach. The segmentation model achieves a dice coefficient of 0.91 and Intersection over Union (IoU) of 0.87, ensuring accurate region extraction. Subsequently, texture features representing thermal distribution patterns are derived using GLCM. These features are then classified using a lightweight ANN designed to balance performance and computational efficiency. Experimental evaluation on both public and self-acquired datasets demonstrates a classification accuracy of 94.11% with low inference latency. The findings indicate that thermography combined with feature-based learning can serve as an effective tool for early-stage screening and decision support, particularly in resource-constrained environments. However, the proposed framework is intended to complement, rather than replace, conventional diagnostic procedures.