Application of Deep Learning in Thyroid Ultrasound Image Analysis

Abstract

Thyroid ultrasound image analysis is critical in detecting and managing thyroid nodules. This thesis explores the application of deep learning models in thyroid ultrasound image analysis, specifically focusing on classification, segmentation, and object detection tasks. Deep learning techniques offer significant advantages for analyzing ultrasound images because they can automatically learn intricate features and handle the complexities inherent in such data. The study employs an extensive dataset of thyroid ultrasound images to evaluate the performance of popular deep learning architectures, including DenseNet, ResNet, and InceptionV3. The findings demonstrate the power of deep learning models in capturing intricate patterns and extracting meaningful features, leading to improved classification performance in thyroid ultrasound image analysis. In addition to classification, the research investigates using self-supervised learning techniques, such as RotNet and denoising autoencoders, for thyroid nodule classification tasks. By leveraging unsupervised feature learning, these techniques reduce reliance on labeled datasets and enhance model generalization. The study explores the potential of self-supervised learning in improving the robustness and adaptability of deep learning models in the context of thyroid ultrasound image analysis. For the segmentation task, the study employs the U-Net architecture, known for its ability to outline the boundaries of objects in medical images accurately. Precise segmentation of thyroid nodules enables accurate measurements and supports informed decision-making. Experimental results demonstrate the effectiveness of the U-Net architecture in achieving specific segmentations for nodules with well-defined boundaries. Furthermore, the study extends to object detection using the Mask R-CNN model. This approach allows for identifying and localizing thyroid nodules within ultrasound images, which may help provide comprehensive assessments and aid treatment planning.