Deep Learning for Ultrasound Classification to Identify Noninvasive Follicular Thyroid Neoplasms with Papillary-Like...

There is substantial demographic and clinical overlap between noninvasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) and other encapsulated tumors. This study aimed to evaluate the feasibility of applying deep learning technology to ultrasound images for the classification and identification of NIFTPs. Preoperative ultrasound images of encapsulated follicular thyroid tumors were divided into a training cohort for training and internal validation, and a prediction cohort for external validation. We used ResNet50 and EfficientNet_B0 models for feature extraction. A total of 279 cases were analyzed, comprising 147 follicular adenomas, 39 follicular thyroid cancers, 47 NIFTPs, and 46 invasive encapsulated follicular variant papillary thyroid cancers. The EfficientNet model demonstrated slightly better classification performance than the ResNet model during internal validation, achieving an accuracy of 0.95 compared to 0.88. Both models exhibited modest performance on a separate prediction cohort, with an accuracy of 0.77. Gradient-weighted class activation mapping (Grad-CAM) indicated that the models primarily focused their attention on the nodule parenchyma. Taken together, despite the overlapping clinical and radiological features of encapsulated follicular thyroid tumors, deep learning models hold the potential to enable preoperative differentiation with promising accuracy, sensitivity, and specificity.