Virginia Via Research Day Book 2026

Undergraduate Student Research Biomedical

04 U-NET BASED AI DATA ANALYSIS OF MOUSE ECHOCARDIOGRAPHY

Junting Zhou, Austin Mason, and Jia-Qiang He, PhD Corresponding author: supremeleader@vt.edu

Department of Biomedical Sciences and Pathobiololgy, College of Veterinary Medicine, Virginia Tech, Blacksburg, Virginia

Cardiovascular disease remains the leading cause of death worldwide. Animal models, especially rodents, are commonly utilized for disease modeling in pre-clinical cardiovascular research. One of the primary methods to evaluate cardiac functions is echocardiography (Echo). Echo is a non-invasive and highly cost-effective approach to be able to record static images and video clips of the heart chamber size, wall thickness, and cardiac movements when its probe is placed on the animal's chest. By analyzing these component’s dimensions, key cardiac output signals, such as ejection fraction (EF%), fraction shortening (FS%), and diastole/systole inner diameters of the ventricular chambers, can be determined. However, most of the Echo data have to be analyzed manually, especially in the rodent study. As a result, the outcomes of the manually-analyzed Echo data are easily influenced by user bias, and potential errors can be occurred in data analysis. Individuals with various levels of experiences in Echo or even with different Echo machines may draw distinct conclusions on the same dataset. In addition, manual analysis is inherently labor-intensive. To overcome this issue, AI based tools, such as PanEcho and EchoNet, have been devised to automate the echo analysis process

in human clinic. These models are typically based on convolutional neural networks (CNN), a class of well-established, computationally efficient, feed forward deep learning models specializing in image recognition. In practice, these models demonstrate high accuracy and are consistent with manual analysis by human experts. Despite these advancements, as of now, almost all rodent Echo data are still being analyzed by individual researchers due to various reasons. Although the general architecture and physiology of the rodent hearts are similar to those of humans, some of the key parameters, such as heart rate, and systolic/diastolic timing, are dramatically different between the two species. This negatively impacts the usability of AI-based human echo tools on the rodent models, especially in regard to B-mode analysis. The development of easily accessible, rodent-specific Echo AI tools is often in an urgent need. Currently, rodent focused solutions, such as the mouse-echocardiography neural net (MENN), employ U-Net-based CNN architectures and demonstrate strong potential, but challenges remain in robustness, cross-machine compatibility, and user accessibility. This project seeks to improve the utility of MENN by addressing two critical limitations. First, robustness will be

enhanced by partially retraining weights of MENN to adapt to situations where rodent Echoes are recorded using a relatively lower-frequency ultrasound probe. Second, compatibility and usability will be improved by modifying the user interface to support a wider range of inputs and allow manual region of interest (ROI) selection. This feature will enable researchers to intervene when automated segmentation fails, ensuring the critical measurements resting physiologically valid. These improvements could potentially accelerate pre-clinical cardiac research, reduce inter-user bias, and strengthen the translational bridge between rodent models and human cardiology.

Table of Contents

125 2026 Research Recognition Day