(C-84) Assessing Cardiac Functions of Zebrafish from Echocardiography Using Deep Learning

Introduction: Cardiovascular disease (CVD) is a significant global health concern, accounting for approximately 32% of all global deaths. Zebrafish (Danio rerio) have emerged as a valuable model organism for cardiovascular research due to their rapid development, ease of maintenance, genetic tractability, and transparent embryos. This study aims to develop automated segmentation tools for zebrafish echo-videos using deep learning techniques to improve the accuracy, efficiency, and objectivity of analyzing complex biomedical data.

Materials and Methods: A dataset of 1005 frames from 164 zebrafish echo-videos was manually annotated using Vevo Lab software and masks representing the ventricular area were generated through ImageJ software. We utilized group k-fold cross-validation (k = 5) in data splitting and our model evaluation. Unsupervised image and video segmentation methods were initially explored but needed to be more robust for echo-video segmentation and ejection fraction assessment. Supervised image segmentation was employed, using deep learning models, including Resnet, Efficientnet, ResNeXt, and Mobilenet as encoders, and Unet and Unet++ as decoders. The training process was executed using NVidia’s A10 and H100 GPU, which was provided by an online cloud service. Performance evaluation was conducted using the dice coefficient (DC) and intersection over union (IoU) metrics.

Results, Conclusions, and Discussions: The Unet++ and Efficientnet architectures outperformed the other encoder–decoder architectures, achieving validation Dice coefficients of 0.967 and validation IoU scores of 0.937. The automatic ejection fraction (EF) calculation method demonstrated high accuracy with an average error rate of 9.83%. This study highlights the potential of the proposed method for the efficiency and accuracy of EF calculation in zebrafish but also reveals the challenges and limitations of using deep learning methodologies on small and domain-specific datasets. Further optimization and improvement of deep learning models will enable accurate and efficient evaluation of cardiovascular metrics from echo-videos, supporting research on cardiovascular development, disease diagnosis, and drug discovery in zebrafish models. Future research should examine more comprehensive and higher-quality datasets to enhance the efficacy of unsupervised learning approaches and explore the combination of supervised and unsupervised learning methods to improve medical image segmentation accuracy.

Acknowledgements: The authors would like to acknowledge the financial support from the NIH SBIR grant #R44OD024874 (H.C.), the NIH HL107304 and HL081753 (X.X.).