A Study on Segmentation Architectures for Thyroid Scintigram Delineation

Abstract

Automated thyroid segmentation in scintigraphic imaging is critical for functional assessment but remains challenging due to low resolution, photon noise, diffuse boundaries, and class imbalance. This study compares four deep learning architectures - U-Net, U-Net++, TransUNet, and Swin-UNet - trained and evaluated on a public thyroid scan dataset (189 training, 23 validation, 23 evaluation studies) under identical conditions. TransUNet achieved the highest accuracy, slightly outperforming U-Net while matching its boundary precision. U-Net followed closely with superior efficiency, lowest latency and highest throughput, making it ideal for realtime deployment. Swin-UNet showed competitive but lower performance on fine structures, while U-Net++ exhibited substantially reduced accuracy and high variability. These results indicate, that increased architectural complexity does not necessarily improve performance. While transformer-based models offer modest gains through global context modeling, convolutional architectures remain robust and efficient. U-Net provides the best accuracy-efficiency trade-off;