CardiSort

CardiSort classifies cardiac magnetic resonance (MR) images using a convolutional neural network to automate sorting by sequence type and imaging plane across multiple vendors.

Key Features:

• Multivendor compatibility: Trained on data from diverse sources including four centers and three different vendors to improve cross-vendor robustness.
• Dual-head architecture: A two-head CNN simultaneously classifies images by sequence type (17 types) and imaging plane (10 planes).
• Single Vendor Training (SVT): Trained on single-center data comprising 234 patients with sequence/plane/overall accuracy of 85.2%/93.2%/81.8% and F1-score 0.82.
• Multivendor Training (MVT): Trained on multicenter data from 434 patients across three centers with sequence/plane/overall accuracy of 96.1%/97.9%/94.3% and F1-score 0.94.
• External validation: MVT was validated on images from three previously unseen magnet systems from two vendors (80 patients) with sequence/plane/combined accuracy of 92.7%/93.0%/86.6% and F1-score 0.86.

Scientific Applications:

• Automated sequence selection for post-processing: Automates the initial sequence selection step in clinical post-processing pipelines for cardiac MR studies.
• Development of fully automated post-processing systems: Serves as a component for building end-to-end automated cardiac MR post-processing workflows.
• Cross-vendor generalizability assessment: Enables evaluation of model performance and generalizability across different magnet systems and vendors.
• Identification of underrepresented classes: Highlights limitations for underrepresented or highly variable sequences (e.g., perfusion imaging) to guide targeted dataset augmentation or refinement.

Methodology:

Retrospective multi-center cardiac MRI datasets were collected and a convolutional neural network with a dual-head architecture was trained and validated using expert radiologist ground-truth labels, including external validation on unseen magnet systems.