Automated assessment of cardiac pathologies on cardiac MRI using T1-mapping and late gadolinium phase sensitive inversion recovery sequences with deep learning

Aleksandra M. Paciorek, Claudio E. von Schacky, Sarah C. Foreman, Felix G. Gassert, Florian T. Gassert, Jan S. Kirschke, Karl-Ludwig Laugwitz, Tobias Geith, Martin Hadamitzky, Jonathan NadjiriBmc Medical Imaging 2024 · DOI

TL;DR

This paper introduces a deep learning model that can automatically detect cardiac pathologies from cardiac MRI scans using T1-mapping and late gadolinium-enhanced (LGE) PSIR sequences. The model shows promising diagnostic performance, with the LGE PSIR-based model achieving 88% accuracy and the T1-mapping model reaching 70%, highlighting the potential for faster and more accessible cardiac diagnostics.

What Problem Does This Paper Address?

Cardiovascular diseases are a leading cause of death and a significant burden on healthcare systems. Early and accurate diagnosis is crucial for effective treatment and improved patient outcomes. Cardiac Magnetic Resonance (CMR) imaging is a powerful tool for assessing heart health, but its interpretation can be time-consuming and requires specialized expertise. This can lead to delays in diagnosis and treatment, especially in high-volume clinical settings or areas with limited access to expert radiologists. The paper aims to address this by developing an automated system using deep learning to assist in the detection of cardiac pathologies from CMR images, thereby streamlining the diagnostic workflow and potentially improving patient care.

Methods

The researchers employed a deep learning approach using a pre-trained DenseNet-161 convolutional neural network (CNN) architecture, implemented with PyTorch and fastai. They trained two separate models: one for T1-mapping sequences and another for late gadolinium phase-sensitive inversion recovery (PSIR) sequences.

Key methodological aspects include:

  • Transfer Learning: The DenseNet-161 model was initialized with weights pre-trained on the ImageNet dataset, a common practice that leverages knowledge learned from a large general image dataset to improve performance on medical imaging tasks, especially when labeled medical data is scarce.
  • Data Augmentation: Techniques like random rotation and mixup were used to artificially increase the size and diversity of the training dataset, enhancing the model's robustness to variations in image orientation and improving its generalization capabilities.
  • Loss Function and Optimizer: Categorical cross-entropy was used as the loss function, and stochastic gradient descent with momentum was employed for weight optimization. A cyclic learning rate was also utilized.
  • Model Architecture Modification: The final classification layer of the DenseNet-161 was adapted from 1000D to 2D to suit the binary classification task (normal vs. abnormal).
  • Visualization: Gradient-weighted class activation maps (Grad-CAMs) were used to visualize which regions of the MRI scans the DL model focused on when making its predictions, providing insights into the model's decision-making process and helping to identify potential biases or areas of focus.

Datasets

The study utilized a dataset of cardiac MRI scans from the German Heart Center Munich.

  • Patient Cohort: The dataset included 200 patients: 137 diagnosed with various cardiac pathologies and 63 classified as normal.
  • Pathologies: The abnormal cases encompassed a diverse range of 16 cardiac pathologies, including acute and chronic myocardial infarction, myocarditis (new and old), dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), and others. This broad spectrum aims to reflect a more realistic clinical scenario compared to studies focusing on a single pathology.
  • Data Split: The dataset was partitioned into training (65%), validation (15%), and hold-out testing (20%) sets, with stratified sampling employed to handle the imbalanced class distribution.
  • Image Sequences: The study used two types of CMR sequences: T1-mapping and late gadolinium phase-sensitive inversion recovery (PSIR).

Key Insights and Ideas

  • Dual-Sequence Approach: Evaluating both T1-mapping and LGE PSIR sequences separately highlights their distinct strengths and weaknesses for automated cardiac pathology detection. T1-mapping's advantage lies in its speed and lack of contrast agent requirement, making it a potentially more accessible screening tool.
  • Real-World Data Complexity: The study's inclusion of a diverse range of pathologies and an unbalanced dataset more closely mimics real-world clinical scenarios, presenting a more challenging yet realistic evaluation of the DL model's performance.
  • Grad-CAMs for Interpretability: The use of Grad-CAMs is crucial for understanding why the model makes certain predictions. It helps confirm that the model is focusing on relevant cardiac structures and pathologies, rather than relying on spurious correlations or artifacts.
  • Trade-off Between Sensitivity and Specificity: The study acknowledges a trade-off, particularly with the PSIR model achieving 100% sensitivity (no missed abnormal cases) at the cost of lower specificity (more false positives). This prioritization is justified by the goal of an inline AI function that flags potential issues for further review, aiming to reduce the risk of missing critical diagnoses.

Key Results

The deep learning models demonstrated promising performance on the test sets:

  • PSIR-based Model: Achieved an accuracy of 88%, with 100% sensitivity, 38% specificity, and a 0% false-negative rate.
  • T1-mapping-based Model: Achieved an accuracy of 70%, with 78% sensitivity, 54% specificity, and an 11% false-negative rate.
  • Grad-CAM Visualization: Confirmed that the DL models primarily focused on the myocardium and areas of cardiac pathology when assessing images, indicating that the learned features were clinically relevant.
  • Comparison to Other Studies: The paper notes that their approach differs from many previous studies by using image-based features from less commonly explored sequences (T1 mapping and LGE PSIR) and handling a broader range of pathologies in an unbalanced dataset.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Binary Classification: The current model performs binary classification (normal vs. abnormal), and future work could extend this to multiclass classification to differentiate between specific pathologies.
  • Dataset Imbalance and Thresholds: The unbalanced dataset and chosen thresholds contributed to the discrepancy between sensitivity and specificity. Future work could involve larger datasets and more refined threshold tuning.
  • Image Quality and Artifacts: Misclassifications can occur due to poor image quality or artifacts. The lower spatial resolution of T1-mapping compared to LGE PSIR might contribute to its lower accuracy.
  • Single Center and Vendor: The study was conducted at a single center using a single MRI vendor. External validation on larger, multi-center, and multi-vendor datasets is necessary to confirm the model's generalizability.
  • Integration with Clinical Data: The model currently relies solely on image data. Incorporating additional patient health records or test results could enhance diagnostic performance.
  • Other MRI Sequences: The applicability of the developed networks to other MRI sequences remains to be explored.
  • Clinical Implementation: A prospective study involving practical clinical implementation on a patient test group is suggested to assess the real-world impact on workflow efficiency and treatment outcomes.

Summary updated

BibTeX
@article{paciorek2024automated,
  author = {Aleksandra M. Paciorek and Claudio E. von Schacky and Sarah C. Foreman and Felix G. Gassert and Florian T. Gassert and Jan S. Kirschke and Karl-Ludwig Laugwitz and Tobias Geith and Martin Hadamitzky and Jonathan Nadjiri},
  title = {{Automated assessment of cardiac pathologies on cardiac MRI using T1-mapping and late gadolinium phase sensitive inversion recovery sequences with deep learning}},
  journal = {Bmc Medical Imaging},
  year = {2024},
  doi = {10.1186/s12880-024-01217-4},
}
Figure