Open-source pre-clinical image segmentation: mouse cardiac magnetic resonance imaging datasets with a deep learning...

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Pre-clinical Research

Background:

• Manual segmentation of cardiac MRI in mice is time-consuming and variable.
• Existing deep learning models do not generalize to pre-clinical cardiac MRI data.
• Lack of public datasets and models hinders large-scale pre-clinical studies.

Purpose of the Study:

• To present the first publicly-available pre-clinical cardiac MRI dataset.
• To develop and release an open-source deep learning segmentation model for mouse cardiac MRI.
• To establish a benchmark for pre-clinical cardiac MRI analysis.

Main Methods:

• Created a dataset of cine short-axis cardiac MRI from 130 mice with diverse phenotypes at 9.4T.
• Included expert manual segmentations for left ventricular blood pool and myocardium.
• Developed an open-source UNet3+ based deep learning model for segmentation.
• Evaluated the model on internal and external test datasets across different field strengths (7T, 9.4T, 11.7T).

Main Results:

• Deep learning model achieved segmentation in ~4.6s per cine stack, over 6,000x faster than manual analysis.
• High segmentation accuracy with Dice scores ≥ 0.91 for blood pool and myocardium.
• Excellent agreement for functional parameters (ICC ≥ 0.89), comparable to human inter-observer variability.

Conclusions:

• The first open-access mouse cardiac MRI dataset and open-source DL model are provided.
• This resource establishes a benchmark for pre-clinical cardiac MRI.
• Enables reproducible, scalable, and community-driven development for cardiovascular research acceleration.