Thyroid texture classification using machine learning in conjunction with autoregressive modeling and deep learning : [kumulative Dissertation]

Abstract

Analysis of medical images play a crucial role in diagnosis and treatment of several diseases in human body. Texture classification is an important tool for segmentation, tissue characterization and organ/boundaries detection in medical images. In this research, we mainly focussed on characterization of tissues in thyroid Ultrasound (US) images so that a thyroid region can be classified from the non-thyroid region. The goal of this thesis was to segment a thyroid region in 2D thyroid US images by characterizing the thyroid and non-thyroid textures using several image and signal based texture classification approaches. The segmented thyroid images could be used for 3D reconstruction and computation of the thyroid volume. The volumetric analysis of thyroid allows for diagnosis of probable thyroid diseases. The first part of the thesis focuses on using classic image based methods to segment the thyroid. Active Contours Without Edges (ACWE), Graph Cut (GC) and Pixel Based Classifier (PBC) were used for thyroid segmentation in 2D US images. These approaches were compared based on accuracy, computation time, robustness and level of human interactions required. The second part explains a novel feature extraction technique that parametrically models a signal version of the US image as a data resulting from a dynamical process. Autoregressive (AR) modelling is used to compute several energy based features which are used to train different machine learning (ML) based classifiers. Similarly, Higher Order Statistical Analysis was also used as another feature extraction technique in a separate study. The extracted features were then used for training several machine learning classifiers. The trained classifiers were later used to classify the thyroid and non-thyroid textures. The final part focuses on using current deep-learning (DL) based approaches to segment the thyroid. We trained a U-Net Convolutional Neural Network and a Fully Connected Convolutional Neural Network (FCNN) using several 2D thyroid US images. The trained CNNs were used for segmenting the test thyroid US images. Throughout the study, we saw that all the ML and DL based approaches require large amount of training images to segment the thyroid with significant accuracy. Hence, we explored the possibilities of generating synthetic 2D thyroid US images using Generative Adversarial Networks (GAN).