Enhancing automatic diagnosis of thyroid nodules from ultrasound scans leveraging deep learning models

Abstract

Abstract The thyroid gland is prone to various diseases, including thyroid nodules. Ultrasound is the primary diagnostic tool, but classification accuracy is often limited by radiologist expertise. Integrating Artificial Intelligence, particularly Deep Learning, offers the potential to enhance diagnostic reliability. This study investigates whether transfer-learning Convolutional Neural Networks (CNNs) can reliably classify TNs using a publicly available, biopsy-verified ultrasound dataset of 483 images (197 benign, 286 malignant). Nine pre-trained CNNs (ResNet50, ResNet101, VGG16, VGG19, DenseNet121, EfficientNetB0, InceptionV3, InceptionResNetV2, and Xception) were evaluated with transfer learning, data augmentation, class balancing, and tenfold cross-validation. ResNet50 achieved the best performance (accuracy 96.90%, Area Under the Receiver Operating Characteristic Curve (AUC) 0.97, precision 96.93%, recall 96.90%, F1-score 96.90%), followed by ResNet101 (94.75% accuracy, AUC 0.95) and EfficientNetB0 (93.09% accuracy, AUC 0.94). Other models achieved accuracies between 87–90% with AUC values of 0.89–0.93. Augmentation and balancing strategies effectively reduced class bias and improved generalization across all models. These findings highlight the superiority of ResNet50 while underscoring the broader potential of CNN-based transfer learning as a reliable decision-support approach for thyroid nodule classification.