A study of machine learning models for rapid intraoperative diagnosis of thyroid nodules for clinical practice in China

Abstract

AbstractBackgroundIn China, rapid intraoperative diagnosis of frozen sections of thyroid nodules is used to guide surgery. However, the lack of subspecialty pathologists and delayed diagnoses are challenges in clinical treatment. This study aimed to develop novel diagnostic approaches to increase diagnostic effectiveness.MethodsArtificial intelligence and machine learning techniques were used to automatically diagnose histopathological slides. AI‐based models were trained with annotations and selected as efficientnetV2‐b0 from multi‐set experiments.ResultsOn 191 test slides, the proposed method predicted benign and malignant categories with a sensitivity of 72.65%, specificity of 100.0%, and AUC of 86.32%. For the subtype diagnosis, the best AUC was 99.46% for medullary thyroid cancer with an average of 237.6 s per slide.ConclusionsWithin our testing dataset, the proposed method accurately diagnosed the thyroid nodules during surgery.