A Multi-Head Attention-Based Lightweight Generative Adversarial Network for Thyroid Ultrasound Video Super-Resolution

Thyroid cancer is a prevalent malignancy, highlighting the critical need for early detection of thyroid nodules. Despite ultrasound being the primary diagnostic modality, its limited resolution often hampers the accuracy of identifying malignant nodules. Method: We present a cutting-edge deep learning approach, the Multi-Head Attention-based Generative Adversarial Network (MHVSR), tailored for the super-resolution of thyroid ultrasound videos. This innovative method encompasses a lightweight optical flow extraction network for swift analysis, a recurrent fusion reconstruction module based on multi-head mechanism, and a hybrid loss function designed to expedite network convergence. Result: The MHVSR showcased exceptional performance on 4× super-resolution task on Thyroid dataset, achieving a PSNR of 31.43dB and an SSIM of 0.98, surpassing current VSR algorithms. It delivered reconstructions that closely approximated the quality of original high-resolution videos, while meeting the stringent real-time diagnostic criteria with an impressive inference speed of 27.89FPS. Furthermore, on the Liver test dataset, MHVSR demonstrated its robustness with a PSNR of 39.39dB and an SSIM of 0.97, outperforming other algorithms. Conclusion: MHVSR not only achieves state-of-the-art super-resolution in ultrasound videos but also satisfies the demanding real-time diagnostic requirements, marking a significant leap forward in the field of medical imaging, particularly for the diagnosis of thyroid nodules.