Automated Thyroid Nodule Characterization from Ultrasound Images Using Deep Learning Architectures

Thyroid nodules are one of the most common diagnoses when undertaking ultrasonography examinations. Despite the fact that most of the thyroid nodules are benign, it is still difficult to distinguish between the two types of nodules. The problem arises mainly because of the great similarity of the nodules in terms of their appearances and variation in sonographic features such as size, contour, echogenicity, internal structure, and calcification of the nodules.

The use of deep learning technologies is an effective way of assisting in diagnosing thyroid nodules since convolutional neural networks are capable of automatically learning discriminative features of images without manually extracting them. In this research, two state-of-the-art architectures of convolutional neural networks, DenseNet-121 and EfficientNet-B0, have been considered for automatic thyroid ultrasound image classification into benign and malignant nodules. These models were developed based on the thyroid ultrasound images taken from the publicly available Digital Database of Thyroid Ultrasound Images dataset.

The comparison of the two model performances was done using some of the model evaluation methods for classification such as accuracy and Area Under the Curve (AUC). From the analysis, both models have performed excellently. Nonetheless, the DenseNet-121 model outperformed the other in that it was able to obtain an accuracy of 99.3% and an AUC of 0.998. From this analysis, one may conclude that the DenseNet-121 model is capable of extracting meaningful clinical information from the ultrasound images of the thyroid gland.