Exploring the effect of domain-specific transfer learning for thyroid nodule classification

Thyroid nodule evaluation using ultrasound is dependent on radiologist experience, but deep learning (DL) models can improve intra-reader agreements. DL model development for medical imaging with small datasets can be challenging. Transfer learning is a technique used in the development of DL models to improve model performance in data-limited scenarios. Here, we investigate the impact of transfer learning with domain-specific RadImageNet dataset and non-medical ImageNet on the robustness of classifying thyroid nodules into benign and malignant. We retrospectively collected 822 ultrasound images of thyroid nodules of patients who underwent fine needle aspiration in our institute. We split our data and used 101 cases in a test set and 721 cases for cross-validation. A Resnet-18 model was trained to classify thyroid nodules into benign and malignant. Then, we trained the same model architecture with transferred weights from ImageNet and RadImageNet. The model without transfer learning for thyroid nodule classification achieved an AUROC of 0.69. The AUROC of our model after transfer learning with ImageNet pre-trained weights was 0.79. Our model achieved an AUROC of 0.83 from transfer learning of the RadImageNet pre-trained weights. The AUROC from the classification model without transfer learning significantly improved after transfer learning with ImageNet (p-value = 0.03) and RadImageNet transfer learning (p-value <0.01). There was a statistically significant distinction in performance between the model utilizing RadImageNet transfer learning and that employing ImageNet transfer learning (p-value <0.01). We demonstrate the potential of RadImageNet as a domain-specific source for transfer learning in thyroid nodule classification.