Leveraging a CMR Foundation Model for Automated Classification of Stress Perfusion CMR Datasets: Initial Results Using the SCMR Registry

Presenting Author(s)

• 

  Zhuoan Li, MSc

  PhD Candidate
  Laboratory for Translational Imaging for Microcirculation, Weldon School of Biomedical Engineering, Purdue University
  Indianapolis, Indiana, United States

Primary Author(s)

• 

  Zhuoan Li, MSc

  PhD Candidate
  Laboratory for Translational Imaging for Microcirculation, Weldon School of Biomedical Engineering, Purdue University
  Indianapolis, Indiana, United States

Co-Author(s)

• 

  M. Berk Sahin

  PhD student
  Purdue University
  West Lafayette, Indiana, United States
• DY

  Dilek M. Yalcinkaya, MSc

  PhD Candidate
  Purdue University
  West Lafayette, Indiana, United States
• AM

  Arian M. Sohi, BSc

  PhD Student
  Laboratory for Translational Imaging for Microcirculation, Weldon School of Biomedical Engineering, Purdue University
  Indianapolis, Indiana, United States
• LZ

  Luis Zamudio, BSc

  Research Engineer
  Purdue University
  Indianapolis, Indiana, United States
• 

  Khalid Youssef, PhD, MSc

  Assistant Professor
  Indiana University, Department of Radiology and Imaging Sciences
  Indianapolis, Indiana, United States
• 

  Michael D. Elliott, MD

  Director of Cardiac MRI
  Atrium Health
  Charlotte, North Carolina, United States
• 

  Venkateshwar Polsani, MD

  Director of Cardiovascular Imaging
  Piedmont Heart Institute
  Atlanta, Georgia, United States
• MT

  Matthew S. Tong, DO

  Associate Professor - Clinical
  The Ohio State University
  Columbus, Ohio, United States
• 

  Dipan J. Shah, MD

  Chief, Division of Cardiovascular Imaging Director, Cardiovascular MRI Laboratory
  Weill Cornell Medical College, Houston Methodist DeBakey Heart and Vascular Center, Houston, Texas, USA
  Houston, Texas, United States
• 

  Orlando P. Simonetti, PhD, FSCMR

  Professor
  The Ohio State University
  Columbus, Ohio, United States
• BS

  Behzad Sharif, PhD

  Associate Professor
  Purdue University
  Indianapolis, Indiana, United States

Background: Stress first-pass perfusion (FPP) CMR is an established tool for assessing myocardial ischemia and ischemic burden, with strong prognostic value in patients with suspected coronary artery disease.^1,2 An automated approach for classification of stress FPP data using artificial intelligence techniques is lacking. Automated classification is especially challenging in heterogeneous multi-center datasets, where established convolutional neural network (CNN)³ or ResNet⁴ architectures may have limited performance. Recent work has shown that Foundation models (FMs) pre-trained on large unlabeled cine CMR datasets can effectively encode/capture cardiac structure and motion, thereby enhancing the performance of downstream tasks including classification.⁵ We hypothesize that adopting a FM trained using cine CMR data and fine-tuning it with stress FPP data enables accurate binary classification of normal vs. abnormal stress FPP cases, supporting rapid automated analysis. We report preliminary results for such a method by leveraging a multi-center dataset derived from the SCMR Registry.⁶

Methods: The FM consisted of a CNN encoder and a transformer encoder initialized with pretrained parameters using a large cine CMR dataset,⁵ followed by fully connected layers for a probability-based output (probability of normal/abnormal class). FPP data were obtained from the SCMR Registry⁶ across 4 centers (n=433 motion-corrected series). The dataset was split into training (n=314), validation (n=68), and testing (n=51). Groundtruth binary labels (normal vs. abnormal) were defined based on summed stress scores ("abnormal" defined as summed score >1). We fine-tuned the cine FM using this multi-center FPP dataset. For comparison, we also trained two other models using established CNN and ResNet architectures. All models were optimized with cross-entropy loss. Performance was evaluated based on area under-the-ROC curve (AUC).

Results: Fig. 2 shows representative cases with the top row showing an abnormal stress FPP series with subendocardial defects. The fine-tuned FM correctly classifies this case as abnormal with high confidence (89.9% probability) while the other two methods fail. The bottom row shows a normal case wherein the FM correctly classifies the case as normal (see caption). Fig. 3 illustrates the overall diagnostic performance, where the proposed fine-tuned FM approach achieved an AUC of 0.860 ± 0.013 (5-fold cross-validation) outperforming both CNN and ResNet (0.832 ± 0.017 and 0.726 ± 0.038, p< 0.01).

Conclusion: Our preliminary results suggest that fine-tuning a CMR foundation model enables automatic binary classification of multi-center stress FPP data. By encoding information from a large cine CMR dataset (n=74,916), the FM may provide high accuracy for this task if it is fine-tuned with a sufficiently large FPP dataset. Automated classification may enable A.I.-assisted analysis of stress CMR, whereby the cases read as normal with high confidence help reduce the workload of expert readers.

Figure 1. Description of the proposed method for automated stress perfusion CMR classification using a pretrained CMR foundation model using cine data (ref. [5]). As shown in the upper panel, this is a self-supervised foundation model trained on a large number of cine CMR studies (n=74,916 patients) using a masked autoencoding strategy. Cine images are randomly masked and passed through a CNN encoder and a Vision transformer encoder-decoder framework. The network is optimized with a mean squared error (MSE) loss to reconstruct missing patches, thereby capturing spatial representations of cardiac structure and motion. The lower panel shows the pipeline of our downstream Stress FPP binary classification task. We used multi-center motion-corrected stress FPP CMR series from the SCMR Registry to fine-tune the foundation model. Perfusion series were first processed by a CNN encoder and subsequently passed through a transformer encoder initialized with pretrained weights. The resulting representations were mapped to class probabilities (normal vs. abnormal) via fully connected linear layers, and the model was optimized using cross-entropy (CE) loss against ground-truth labels.
Figure 2. Representative stress FPP cases from a four-center dataset derived from the SCMR Registry with binary classification (normal vs. abnormal) using the proposed fine-tuned foundation model (FM) approach vs. alternative deep-learning-based approaches. Two representative stress FPP CMR series are shown with corresponding 17-segment bullseye maps providing ground-truth perfusion scores. Binary classification was defined by the summed stress score (summed score >1 considered abnormal). For each case, predicted class probabilities from the proposed fine-tuned foundation model and the two alternative methods (CNN and ResNet) are reported on the right-handed side. Top row: an abnormal case with subendocardial perfusion defects (green arrows) scored as abnormal as seen in the bull's eye map (scores from the SCMR registry). Our proposed fine-tuned FM approach correctly classified it as abnormal with high confidence (89.9%), whereas CNN and ResNet classifiers misclassified it as normal. Bottom row: A normal case correctly identified by our model as normal (82.0% probability) and misclassified by CNN and ResNet.
Figure 3. Comparison of the diagnostic performance of the proposed fine-tuned FM approach versus the two alternative deep learning-based methods (CNN and ResNet) using receiver operating characteristic (ROC) curve and 5-fold cross-validation. The multi-center dataset was derived from the SCMR Registry. Across folds, the proposed method achieved the highest overall AUC (0.860 ± 0.013), outperforming CNN (0.832 ± 0.017) and ResNet (0.726 ± 0.038) with p<0.01 for both comparisons.