RED-Net: A Neural Network for 3D Thyroid Segmentation in Chest CT Using Residual and Dilated Convolutions for Measuring Thyroid Volume

Unlike the lungs or the heart, the thyroid gland is not a primary target in chest computed tomography (CT) scans and is relatively small; hence, it is difficult for radiologists to always clinically delineate it in chest CT to incidentally detect a goiter. We designed a residual and dilated convolution neural network (RED-Net), which automatically measures thyroid volume by segmenting the thyroid gland in contrast-enhanced chest CT scans. Its fundamental structure comprises a residual downsampling and upsampling pathway, complemented by a parallel dilated convolution module. This combination allows the model to extract features at multiple scales and capture contextual information to effectively segment even tiny thyroid glands in the complex anatomical structures observed in chest CT scans. Additionally, we constructed training and validation sets comprising CT scans of 1,150 adults (aged <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\ge 19$ </tex-math></inline-formula> years) who underwent chest CT scans at the National Cancer Center and included data of those without a history of thyroid nodules, C73 diagnosis, or thyroid surgery before scanning procedure. We evaluated the performance of our method on a test dataset (600 patients) comprising chest CT scans of individuals collected at Chungbuk National University Hospital using the same criteria. The results showed that it achieved state-of-the-art performance with a Dice similarity coefficient of 0.8901.