Localization with deep learning networks for super-resolution ultrasound imaging

Abstract

OBJECTIVE: The objective was to optimize an acquisition methodology and deep learning network to localize a microbubble (MB) contrast agent on a preclinical ultrasound (US) platform with the goal of moving super-resolution ultrasound (SRUS) imaging towards clinical adoption. METHODS: A deep learning network was optimized based on the latest advances in computer vision, convolutional neural networks, and transformer architectures. Synthetic data were produced in an US simulation of tissue with MB at various concentrations flowing in a vascular model. The network was programmed in PyTorch and trained on 12 000 synthetic images. US data were collected with a Vevo 3100 (FUJIFILM VisualSonics Inc) using an MX201 linear transducer. In vivo testing images were obtained from hepatocellular carcinoma (HCC) rat liver tumors. After MB injection (Definity), 3000 frames were collected at 90 Hz. RESULTS: The SRUS in vivo results reveal a high level of detail. The smallest vessel measured 34 μm in diameter. Compared to conventional methods, the network improved performance by 10x on a CPU. A GPU platform could give an additional boost, by as much 100x. CONCLUSIONS: Deep networks for localization show potential for improving performance of SR-US towards a real-time imaging modality. The use of a pre-clinical focused US platform demonstrates that localization can improve the visualization of vascular detail and aid clinical understanding.