MUST 2024

Neural Network-Based Data Recovery Using Spiral Sub-Sampling for Ultrasound Computed Tomography

Jun 12, 2024, 3:00 PM

20m

Eulerzaal (Centrum Wiskunde & Informatica)

oral presentations Machine Learning

Speaker

Hantao Yang (Department of Biomedical Engineering, Huazhong University of Science and Technology)

Description

Ultrasound computed tomography (USCT) has the potential for clinical applications due to its standardized operations and multi-modality. However, obtaining high-quality images requires a complete dataset including all transmitting-receiving pairs, resulting in a time-consuming scanning process and substantial data-processing demands. The limitation restricts the clinical applications of USCT. Reconstructing images directly from sub-sampled radio-frequency (RF) data leads to diminished image quality. To address this issue, we propose an efficient spiral sub-sampling strategy for sparse data acquisition and a data recovery approach based on convolutional neural network (CNN).

The spiral sub-sampling strategy selects receiving channels symmetrically, centering around the transmitting array. Specifically, for the n-th transmitting event, echo data are collected from the n, n+m, n+2m, and subsequent elements. The uniform sub-sampling strategy employs fixed channels for reception. Specifically, for every transmitting event, echo data are collected from the m, 2m, 3m, and subsequent elements. To validate the approach, six sets of complete RF data from the human leg were acquired, the first five sets were used to train the CNN, and the remaining was used for testing. We conducted experiments using the CNN to recover data sub-sampled by a factor of 4× or 8×. To quantitatively show the advantages of the proposed method, we used structure similarity (SSIM) and peak signal-to-noise ratio (PSNR) as metrics to assess the quality of reconstructed images.

In 4× Rx sub-samping experiments, the uniform and spiral methods achieved SSIM and PSNR values of 0.07, 0.08, and 7.16 dB, 6.52 dB higher than the input, respectively. In 8× Rx sub-samping experiments, the uniform and spiral methods achieved SSIM and PSNR values of 0.02, 0.07, and 0.46 dB, 5.88 dB higher than the input, respectively. The results show that for higher downsampling rates, our proposed method demonstrates the potential to utilize the data redundancy in the transmit-receive plane.