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

backgroundLongitudinal cardiovascular magnetic resonance (CMR) in small animal models is essential for studying cardiovascular disease mechanisms and assessing new therapies. However, conventional manual segmentation of ventricular structures from cine short-axis images is slow, labor-intensive, and subject to substantial observer variability, limiting reproducibility in large-scale pre-clinical studies. While deep learning (DL) segmentation has become standard in human CMR, existing models do not generalize to pre-clinical datasets. Current mouse-specific approaches have been constrained by small private cohorts, where the DL segmentation models are not published.

methodsHere, we present the first publicly available pre-clinical CMR dataset, along with an open-source DL segmentation model and a web-based interface for easy deployment. The dataset comprises complete cine short-axis CMR images from 130 mice with diverse phenotypes, acquired at 9.4T. The dataset also contains expert manual segmentations of left ventricular (LV) blood pool and myocardium at end-diastole, end-systole, as well as additional timeframes with artifacts to improve robustness. Using this resource, we developed an open-source DL segmentation model based on the UNet3+ architecture. The model was evaluated against an independent internal test dataset (9.4T, n = 25 mice), as well as two external test datasets (n = 15 mice at 7T, and n = 10 mice at 11.7T) and compared to expert manual segmentations in terms of Dice and functional parameters.

resultsAt inference, the DL segmentation model took <20 ms per 2D image, enabling complete cine stack segmentation in ∼4.6 s-over 6,000-fold speed-up over manual analysis. It achieved high segmentation accuracy in both the blood pool and myocardium in all test datasets (overall Dice ≥ 0.91). In addition, functional parameters showed excellent agreement with manual ground-truth measurements (intraclass correlation coefficient (ICC) ≥ 0.89 for the internal test dataset, and ICC > 0.85 for the external test datasets), with ICC values comparable to human inter-observer variability.

conclusionThis work provides the first open-access mouse cardiac MRI dataset, an open-source DL segmentation model, and an accessible inference platform. These tools establish a benchmark for pre-clinical CMR, enabling reproducible, scalable, and community-driven development of DL methods to accelerate cardiovascular research.