Automated characterization of abdominal MRI exams using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Magnetic resonance imaging (MRI) data is complex and heterogeneous, hindering large-scale multi-institutional studies.
• Standardized tools are needed to automatically identify and characterize key imaging attributes for harmonized data.
• Machine learning models require standardized data for robust training and reliable outcomes.

Purpose of the Study:

• To develop and validate convolutional neural networks (CNNs) for automatic classification of abdominal MRI attributes.
• To classify pulse sequence type, imaging orientation, and contrast enhancement status using distinct CNN models.
• To assess the generalizability and performance of these CNNs on external datasets.

Main Methods:

• Developed three distinct CNNs with similar architectures to classify single MRI slices.
• Trained models to identify 12 pulse sequences, 4 orientations, and 2 contrast classes.
• Employed a majority voting approach for slice-level aggregation and applied Grad-CAM for visualization.

Main Results:

• Achieved high slice-level classification accuracies: 99.51% (pulse sequence), 99.87% (orientation), and 99.99% (contrast).
• Reached 100% volume-level accuracy for all classification tasks using majority voting.
• Demonstrated strong generalizability with >96.9% volume-level accuracy on the Duke Liver Dataset.

Conclusions:

• CNNs can accurately and automatically classify core abdominal MRI attributes.
• Standardized attribute classification enhances MRI data harmonization for machine learning.
• These tools hold significant potential for improving large-scale medical imaging research and clinical applications.