The AI model achieved high Dice scores (up to 0.965) for cardiac MRI segmentation and mean errors of -2 to -3% for LV and RV ejection fraction estimation.
Does the CMR-Seg AI framework accurately perform automated segmentation and quantification of cardiac MRI compared to expert physician measurements?
An open-source, fully automated AI framework (CMR-Seg) provides highly accurate segmentation and functional quantification across diverse cardiac MRI sequences and views.
Absolute Event Rate: 0% vs 0%
Abstract Background Cardiac magnetic resonance (CMR) imaging includes multiple sequences and views, making quantification and analysis a time-intensive process that requires specialized expertise and is prone to inter- and intra-observer variability. Current available commercial tools used in clinical practice are typically designed for a single sequence, view, or label, and often demand significant manual modification of segmentation to achieve accurate quantification. Purpose In this study, we developed and validated an open source fully automated artificial intelligence (AI)-based unified framework for comprehensive CMR analysis. Methods This study included 1000 patients with diverse CMR images from eight different centres, encompassing various scanner models, and imaging settings. For cine imaging CMR scans (short-axis and 4-chamber views) along with corresponding segmentations of the left ventricular myocardium (LVM) and the left and right ventricles (LV and RV, respectively), were collected for model development. The remaining images from two local centers, which included cardiac volumetric and functional measurements by expert physicians, were utilized to evaluate the model's performance. For 2- and 3-chamber CMR (LV and LVM), as well as LGE (short axis, 2,3 and 4-Chamber) images (LV, RV, and LVM), and fat segmentation in 4-chamber images (separately for RV, RA, LV, and LA, CMR and LGE), a human-in-the-loop active learning approach was implemented to refine and enhance the model’s accuracy. A modified UNet network with deep supervision from the nnUZoo framework, enhanced by various image augmentation techniques, was developed for automated segmentation across different imaging modalities. Additionally, cardiac functional metrics were computed from short-axis cine images for quantitative analysis. Results In the evaluation set, the model achieved average Dice scores of 0.930, 0.903, and 0.959 for the RV, LV, and LVM in the short-axis, 0.933, 0.901, and 0.965 for the chamber cine, and 0.889, 0.876, and 0.948 for the short axis LGE images. For fat segmentation in 4-chamber images, an average Dice score of 0.589 was obtained across different chambers (RV: 0.686, RA: 0.438, LV: 0.661, and LA: 0.572). Additionally, the model demonstrated high accuracy in functional assessment, with mean percent errors of -2 to -3% for LV and RV ejection fraction estimation in short-axis cine images. Conclusion We developed and validated an open-source, fully automated artificial intelligence (AI)-based unified framework for comprehensive CMR analysis. This framework enables the segmentation and quantification of various volumetric and functional parameters across different CMR sequences and views. By providing a robust and accessible tool, it lays the foundation for advanced quantification techniques, including radiomics, deep learning, strain analysis, and chamber/fat volumetry, across diverse CMR image types.
Kazaj et al. (Sat,) reported a other. The AI model achieved high Dice scores (up to 0.965) for cardiac MRI segmentation and mean errors of -2 to -3% for LV and RV ejection fraction estimation.