Thyroid Nodule Image Joint Segmentation and Classification Based on Deep Learning

Abstract

A thyroid nodule is a thyroid gland condition that must be diagnosed and treated as soon as possible to mitigate the threat of death in a possible patient. This research proposed a joint segmentation and classification system to detect and identify thyroid nodules in ultrasound images automatically. The proposed scheme as it is envisioned has consist of two stages: in the first stage thyroid image features were calculated via a concatenation of features vector (to enhance the diversity of image features)generated by deep learning-based VGG-19 model in the first place and the second place by deriving handmade features from VGG-SegNet image segmentation model followed by computing a Fuzzy Grey Level Co-Occurrence Matrix (FGLCM) for each segmented image, which has the effect of eliminating directional difference by multi-angle fusing the grey level co-occurrence matrix (GLCM) and calculation of the membership of each pixel to the texture unit after applying the fuzzy c-means algorithm to the grey level co-occurrence matrix. The second stage then involves thyroid image classification based on four types of machine learning techniques namely (Naïve Bayes NB, Decision Tree DT, K-Nearest Neighbor KNN, and SVM-RBF Support Vector Machine based Radial Basis Function). The proposed model has been evaluated based on Thyroid Digital Image Database (TDID), which is a public dataset for thyroid nodule segmentation created by Universidad Nacional de Colombia. The experimental results revealed that the SVM-RBF classifier has achieved a validation accuracy of 99.25% with concatenated features vector.

Keywords

• Thyroid nodule; VGG-SegNet; Deep Learning; VGG 19; Fuzzy co-occurrence matrix.