A Hybrid Deep Neural Network Architecture RefineUnet for Ultrasound Thyroid Nodules Segmentation
Kamar Bouhdiba, Lila Meddeber, Mohammed Meddeber, Tarik Zouagui · 2024
Thyroid nodules are prevalent in the human population, often identified during routine ultrasound examinations. Accurate segmentation of these nodules is crucial for diagnosing and managing potential thyroid malignancies. However, precise segmentation of the complex thyroid tissue remains challenging due to the ultrasound image's low contrast and high noise. Several traditional segmentation approaches to detect thyroid nodules suffer from imprecise localization issues, while deep learning techniques offer promising solutions. This study proposes a new deep hybrid architecture that combines the encoder-decoder structure of U-Net with the multi-resolution RefineNet modules to improve the model's ability to recognize thyroid nodules. The proposed approach demonstrates the impact of combining residual blocks and chained residual pooling (CRP) blocks to effectively capture both local detail and global context, resulting in robust semantic segmentation. This is particularly important in thyroid nodule segmentation, where accurate diagnosis and treatment planning require both fine-scale features and contextual information. The model was trained and verified using the public TN3K dataset, which consists of 3493 thyroid nodule images. Our approach was found to significantly improve the accuracy of semantic segmentation, achieving 94.18% with an IOU of 96.40% compared to state-of-the-art methods.