Abstract
INTRODUCTION: Thyroid nodules are common, and investigation is crucial for excluding malignancy. Increased intranodular vascularity is frequently observed in malignant tumors, which can be detected through increased skin surface temperatures using noninvasive infrared thermography. We aimed to develop a diagnostic tool for thyroid cancer using infrared thermal images combined with an artificial intelligence (AI) algorithm. METHODS: We conducted a prospective cross-sectional study involving participants with thyroid nodules undergoing thyroid surgery. Infrared thermal images were collected using a thermal camera on the day prior to surgery. In combination with the final thyroid pathological reports, we utilized a machine learning model based on the pre-trained ResNet50V2 model, a convolutional neural network, to evaluate diagnostic accuracy for malignancy diagnosis. RESULTS: The study included 98 participants, 58 with malignant thyroid nodules and 40 with benign thyroid nodules, as determined by pathological results. The AI-enhanced infrared thermal image analyses demonstrated good performance in distinguishing between benign and malignant thyroid nodules, achieving an accuracy of 75% and a sensitivity of 78%. These parameters were slightly lower than those of the AI-model predictor that integrated current practice using preoperative thyroid ultrasound findings and cytological results, yielding an accuracy of 81% and a sensitivity of 84%. CONCLUSIONS: The infrared thermal images, assisted by an AI model, exhibit good performance in distinguishing thyroid malignancy from benign nodules. This imaging modality has great potential to be used as a noninvasive screening tool for adjunct evaluation of thyroid nodules.