Improved VGG-16 for Classifying Thyroid Nodule on Thyroid Ultrasound Images

Thyroid nodule composition is one of the determinants of thyroid cancer malignancy. Nodules in the thyroid can be detected by ultrasonography, which is one of the most sensitive imaging methods. However, imaging methods by ultrasonography are susceptible to doctors’ experiences, levels, and other factors. Therefore, a more objective diagnostic system is intended to assist doctors in creating the decision. This study developed a method to help experts define the composition characteristics. The experts have already cropped the dataset and then moved to preprocess using an adaptive median filter. Subsequently, the data were classified with the improved VGG16 into four categories those are cystic, solid, complex, and spongiform. The testing result procured 99.65% for accuracy, 99.98% for the micro area under the curve, and 99.99% for the macro area under the curve. These results indicate that our proposed method can be used in a small dataset to help doctors or experts identify the nodule’s characteristics.