Rapid Fire Abstracts
Martin Hadamitzky, MD
Director
German Heart Center Munich, Germany
Keno K. Bressem, MD
Consultant
German Heart Center Munich, Germany
Celine Dürner, MD
Student
German Heart Center Munich, Germany
Eva Hendrich, MD
Radiologist
Clinic of Radiology, German Heart Center Munich, Technical University of Munich, Germany, Germany
Rafael Adolf, MD
Consultant
German Heart Center Munich, Germany
Era Stambollxhiu, MD
Resident
German Heart Center Munich, Germany
Miriam Kumpf, MD
Resident
German Heart Center Munich, Germany
Alessa Chami, MD
Radiologist
German Heart Center Munich, Germany
Cardiac magnetic resonance imaging (MRI) plays a crucial role in diagnosing various heart conditions. However, manual segmentation and interpretation of multi-sequence MRI data are time-consuming and subject to inter-observer variability. This study aims to develop and evaluate an artificial intelligence (AI) approach for automated myocardial segmentation and subsequent cardiac disease classification using multiple MRI sequences.
Methods:
We developed a two-stage AI pipeline for myocardial segmentation and cardiac disease classification. For segmentation, we trained a 2D-U-Net with EfficientNet-B5 backbone slice-wise on four MRI sequences: T2-weighted, T2 mapping, late gadolinium enhancement (LGE) gradient recalled echo (GRE), and pre-contrast T1 mapping. The segmentation model was trained, validated, and tested on a dataset of 935 cardiac MRI examinations with the data split provided in Table 1.
For disease classification, we trained a 3D ResNet-34 ensemble (one model per sequence type) using the segmented myocardium as input. The model was designed to classify examinations into 12 categories: old myocarditis, dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), non-specific changes, old infarction, normal findings, acute myocarditis, specific cardiomyopathy, acute infarction, amyloidosis, pericarditis, and sarcoidosis. The classification model was evaluated using the same dataset split. The pipeline is illustrated in Figure 1.
Results:
The segmentation model achieved a mean Dice coefficient of 0.89 across all sequences, with individual sequence performance as follows: T2-weighted (0.89), T2 mapping (0.91), LGE GRE (0.87), and pre-contrast T1 mapping (0.90).
With a mean Area Under the Receiver Operating Characteristic Curve (AUC) of 0.75, the classification model demonstrated varying performance across different cardiac conditions, as shown in Figure 2. The model showed the highest performance in identifying amyloidosis (AUC 0.98), pericarditis (AUC 0.92), HCM (AUC 0.92), and DCM (AUC 0.91). Other cardiac conditions proved more challenging to differentiate, with AUCs ranging from 0.51 for acute myocardial infarction to 0.82 for old myocardial infarction.
Conclusion: Our AI-based pipeline demonstrates promising results in automated myocardial segmentation across multiple MRI sequences and shows potential for classifying certain cardiac diseases, particularly cardiomyopathies and amyloidosis. However, the model's performance varies significantly across different cardiac conditions, indicating the need for further refinement. Future work should focus on improving the classification of less distinct cardiac pathologies.
