سحر فتاحي

The thyroid is a butterfly-shaped essential organ located in the front of the neck. When some thyroid cells grow abnormally, protrusions are created on the thyroid. These cells come into the thyroid gland as a mass or nodule called thyroid nodules. Although most thyroid nodules are benign, some of them may contain cancer cells. Ultrasound imaging is the best method for identifying thyroid nodules due to its cost-effectiveness and safety. To diagnose thyroid disorders, it is crucial to separate thyroid nodules from the thyroid gland. Therefore, thyroid segmentation is a necessary step in diagnosing thyroid diseases. With recent advances in artificial intelligence, diagnostic systems can now help specialists in accurately diagnosing thyroid diseases in the early stages while improving accuracy, saving time, and aiding in accurate diagnosis. Machine learning-based methods, especially deep learning methods, can significantly enhance segmentation performance, making them preferable choices for thyroid nodule segmentation. The strength of deep neural network algorithms lies in detecting complex patterns. The aim of this thesis is to design a system using deep learning for automatic thyroid nodule segmentation in medical ultrasound images. Introducing a suitable framework based on CNNs, this thesis presents the Dilated Attention U-Net model that combines Dilated Attention blocks with the U-Net architecture for precise and fast medical image segmentation. All implemented models are eva‎luated and compared using common eva‎luation metrics. By integrating dilated convolutions and attention mechanisms in the U-Net architecture, this network achieved an accuracy of 94.34%. This new network can extend the receptive field while preserving spatial resolution effectively. It can selectively focus on important information features, improving performance in semantic segmentation tasks. Segmentation with a deep learning approach provides good performance but requires a more specific database. Additionally, the advantages and disadvantages of this method in two-dimensional ultrasound imaging have been discussed, which will be useful for future studies such as identifying specific thyroid diseases.

ندول تيروئيد , تصاوير سونوگرافي , قطعه‌بندي , يادگيري عميق

Thyroid nodules , Ultrasound images , Segmentation , Deep learning

sahar fattahi

dr. hamid behnam