An Improved V-Net Model for Thyroid Nodule Segmentation

Early diagnosis of increasingly common thyroid nodules is crucial for effectively and accurately managing the disease’s monitoring and treatment process. In practice, manual segmentation methods based on ultrasound images are widely used; however, owing to the limitations arising from the imaging sources and differences based on radiologist opinions, their standalone use may not be sufficient for thyroid nodule segmentation. Therefore, there is a growing focus on developing automatic diagnostic approaches to assist radiologists in nodule diagnosis. Although current approaches have yielded successful results, more research is needed for nodule detection because of the complexity of the thyroid region, irregular tissues, and blurred boundaries. This study proposes an improved V-Net segmentation model based on fully convolutional neural networks (V-Net) and squeeze-and-excitation (SE) mechanisms for detecting thyroid nodules in two-dimensional image data. In addition to the strengths of the V-Net approach in the proposed model, a squeeze-and-excitation (SE) mechanism was used to emphasize important features and suppress irrelevant features by assigning weights to the significant features of the model. Experimental studies utilized the Digital Database Thyroid Image (DDTI) and Thyroid Nodule 3493 (TN3K) datasets, and the improved V-Net-based model was validated using the V-Net, fusion V-Net, and SEV-Net methods. The results obtained from the experimental studies demonstrate that the proposed model outperforms the V-Net, fusion V-Net, and SEV-Net models, with a Dice score of 84.51% and an IoU score of 76.27% for the DDTI dataset. Similarly, on the TN3K dataset, it achieved superior performance compared to all benchmarked models, with Dice and IoU scores of 83.88% and 75.50%, respectively. When considering the results in the context of the literature, the proposed model demonstrated the best performance among all models, achieving an average score of 80.39% for the DDTI dataset and 79.69% for the TN3K dataset, according to both Dice and IoU metrics. The model, with a Dice score of 84.51%, competes at a competitive level with Ska-Net, which exhibits the best performance in this metric with a score of 84.98% on the DDTI dataset, whereas it achieved the best performance among existing models with an IoU score of 75.5% on the TN3K dataset. The achievement of the proposed model may make it an effective tool that radiologists can use for thyroid nodule detection.