A Systematic Approach to the Detection, Quantification and Classification of Thyroid Nodules in Ultrasound Images using Image Computing, Machine and Deep Learning for Reduced Subjectivity and Inter- and Intraobserver Variability

Abstract

Ultrasound imaging is used as a first and most frequent mode for assessing thyroid nodules. Nodule features like irregular shape, microcalcifications, and taller-than-wide morphology raise suspicion for malignancy. However, classification of thyroid nodules using ultrasound images depends heavily on the interpreting physician's experience and skill, leading to inherent issues with subjectivity and both interobserver and intraobserver variability. This thesis presents a systematic approach for the detection, region estimation and classification of thyroid nodules using ultrasound images aimed at reducing overall subjectivity and inter- and intraobserver variability. This is achieved through the use of texture analysis, feature extraction, machine learning and deep learning using ultrasound images with thyroid nodules. The study begins by differentiating textures of the thyroid gland from surrounding organs in US images using autoregressive features and machine learning. This is followed by the comparison of performances from four different deep learning algorithms for the detection and localization of thyroid nodules. Subsequently, we estimate and quantify the solid and cystic regions within thyroid nodules using textural analysis and machine learning. Lastly, extraction of geometric and morphological features, helps classify thyroid nodules using machine learning techniques that consider the visual characteristics analyzed by physicians according to TIRADS. Thus, providing them with quantifiable evidence that supports the classification process. Future research will focus on refining and validating these models, and determine how best to incorporate them into existing clinical workflows.