Super-resolution Ultrasound imaging via Unpaired Training with the Model-Informed CycleGAN Algorithm

Vassili Pustovalov (1) , Duong-Hung Pham (1) , Denis Kouamé (1)

1 IRIT-MINDS - CoMputational imagINg anD viSion

In traditional ultrasound (US) imaging, there has long been a trade-off between spatial resolution and imaging frame rate. Super-resolution (SR) imaging techniques stand out as extensively studied methods to overcome the spatial resolution limitation. Recently, these techniques have garnered significant attention in ultrafast US imaging research, enabling enhanced visualization of microvasculature. However, current SR methods encounter challenges, including the need to acquire a substantial number of image frames over an extended acquisition time, coupled with intricate post-processing steps. Convolutional neural networks (CNNs) emerge as promising approaches, surpassing classical model-based methods. Nevertheless, they face challenges arising from the lack of in vivo data and the absence of ground truth. In this work, we introduce a model-informed CNN for computing SR images, addressing generalization issues through an unpaired and unlabeled training approach. Our results are compared with those obtained from state-of-the-art techniques on both simulated and in vivo data.

Mots clés: Super Resolution; Ultrasound Imaging; Microvasculature; CNN; Unpaired; Cyclic GAN

Domaines: Imagerie médicale; Son [cs.SD]

Vassili Pustovalov, Duong-Hung Pham, Denis Kouamé. Super-resolution Ultrasound imaging via Unpaired Training with the Model-Informed CycleGAN Algorithm. 21st IEEE International Symposium on Biomedical Imaging (ISBI 2024), May 2024, Athènes, Greece. à paraître. ⟨hal-04493922⟩