Utilizing Eigen Images for Enhanced Microbubble Signal Extraction in Ultrasound Imaging

Introduction: In ultrasound imaging, blood signal is masked by clutter and noise. Singular value decomposition (SVD) can decompose different types of signals depending on their energy represented by singular values. To visualize microvasculature, contrast-enhanced ultrasound imaging using microbubbles (MB) is widely used. In in vivo mouse contrast-enhanced ultrasound imaging, it is important to exclusively extract MB signals from mouse movement, tissue scattering, and clutter. To achieve this, we developed an approach to classify MB signals, mouse movement, tissue back ground signals, and depth-dependent noise by analyzing the left singular matrices of SVD. Because the column space of these left singular matrices represent eigen-images of acquired images, we identified different modes of acquired signals that represent mouse movement, MB signals and noises. We compared the proposed approach with elbow point and hard thresholding [1] methods.

Materials and

Methods: We injected 150 μL (0.75 – 2.0 x 10^9 MBs/mL) MBs into a tumor-bearing B6 mouse through the retro-orbital route. The center frequency of the linear array was 28 MHz, with 128 elements. Pulse-echo radiofrequency (RF) data were acquired using the Vantage 256 system using L35-16vx linear array transducers. We acquired 1000 frames in 2 seconds. Saved RF data were processed using MATLAB. We implemented elbow point method to determine transition between mouse movement, clutter signals, and MB signals. Hard thresholding was mathematically proven to pinpoint the optimum singular values between desired signals and noise. The proposed eigen-image based method compared modes of images of each column in the left singular matrix to identify mouse movement, MB signal and background noise. We analyzed the eigen-image matrix by calculating the absolute difference between each image and its predecessor. We then approximated the resulting plot of frame-to-frame differences using linear regression. This allowed us to identify any underlying trends in how the difference changes throughout the entire dataset. By exploiting these patterns, we isolated and extracted signals indicative of blood motion. The vessel density is defined as the ratio of the number of vessel pixels to the total number of pixels in the tumor region in the reconstructed super-resolution ultrasound images [2].

Results, Conclusions, and Discussions: The elbow point method lacks a specific threshold criterion which might lead to an ambiguous decision boundary. The hard thresholding method requires multiple processing steps to extract MB signal. We performed a two-step process to extract clutter signals, mouse movement, tissue scattering, and depth-dependent noise. The eigen image-based method analyzed differences between consecutive frames to simultaneously identify sections with high variation due to mouse movement, tissue scattering, and depth-dependent noise. The frames with MB signals were extracted by excluding two groups of frames that were determined by the eigen image-based method. Finally, the extracted blood signals from each of the three methods were further processed to generate super-resolution ultrasound images, as shown in figures 1A-C, to visualize microvasculature. The super-resolution ultrasound image reconstructed by the eigen image-based method detects more vessels within the mouse tumor region (Fig. 1D). The eigen image-based method proves to be the most effective approach for extracting the blood signal and generating high-density super-resolution ultrasound images of the blood vessel by detecting MB signals more efficiently compared to widely the used elbow point method and the newly developed hard thresholding method. This approach offers promising implications for enhancing the accuracy and efficiency of ultrasound imaging techniques in biomedical applications.

Acknowledgements (Optional): This work is partially supported by the National Institute of Health (NIH) grant No. GM152704 and the National Science Foundation (NSF) grant No. CBET1943852.