Thyroid Nodules Categorization Based On Margin Features Using Deep Learning

Abstract

Thyroid cancer is a rare malignancy originated in the thyroid gland. It is often found incidentally in patients with thyroid nodules. One of the ways to evaluate its malignancy is by using ultrasonography of thyroid and neck. Many studies conducted experiments around thyroid nodules but none were focusing on thyroid nodules characteristics especially margin characteristic. This characteristic tends to be visibly obvious and can be the lead to another characteristics. We proposed a method to help experts in identifying the regular and irregular categories from margin characteristic. The proposed method consists of pre-processing, segmentation, feature extraction by using nine geometric features, data balancing by using synthetic minority oversampling technique (SMOTE), and classification by using deep learning method. For the training process we used the dataset collected from the Department of Radiology RSUP Dr. Sardjito Yogyakarta Indonesia. We sucessfully obtained accuracy of 94.79%, sensitivity of 94.9%, specificity of 94.1%, PPV of 98.68%, NPV of 80%, F-measure of 94.6% and ROC of 96%.