Rapid Fire Abstracts
Arian M. Sohi, BSc
PhD Student
Laboratory for Translational Imaging for Microcirculation, Weldon School of Biomedical Engineering, Purdue University
Dilek M. Yalcinkaya, MSc
PhD Candidate
Purdue University
Arian M. Sohi, BSc
PhD Student
Laboratory for Translational Imaging for Microcirculation, Weldon School of Biomedical Engineering, Purdue University
Khalid Youssef, PhD
Assistant Research Professor
Indiana University, Department of Radiology and Imaging Sciences
Luis F. Zamudio Rivero, MSc
Research Associate
Laboratory for Translational Imaging of Microcirculation, Krannert Cardiovascular Research Center, Indiana University School of Medicine, Indianapolis, IN
Michael D. Elliott, MD
Director of Cardiac MRI
Atrium Health
Venkateshwar Polsani, MD
Director of CardioVascular MRI and CTA
Piedmont Heart Institute
Rohan Dharmakumar, PhD
Executive Director
Indiana University School of Medicine
ROBERT M. JUDD, PhD
Professor Emeritus
Duke University
Duke University
Matthew S. Tong, DO
Associate Professor - Clinical
The Ohio State University
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
Orlando P. Simonetti, PhD
Professor
The Ohio State University
Behzad Sharif, PhD
Associate Professor of Biomedical Engineering
Purdue University
Accurate segmentation of stress first-pass perfusion (FPP) CMR is critical for reliable myocardial blood flow analysis. Automated artificial intelligence (AI) solutions are essential for overcoming the inefficiencies and inconsistencies inherent in manual contouring of stress/rest FPP studies [1,2]. Recent advances in deep neural network (DNN)-based segmentation of CMR datasets show that uncertainty-guided model selection improves generalization to external datasets [3,4]. In this work, we propose a new approach that combines state-of-the-art (SOTA) uncertainty-based DNN model selection with CMR physics-informed features to improve segmentation accuracy across multi-center stress FPP datasets.
Methods:
The DNN models in this study were trained using a motion-corrected stress FPP internal dataset (n=80 stress/rest studies) and tested on three external sites from the SCMR Registry [5]. As described in Fig 1, in the first step, we pre-select the top 10 models based on the lowest uncertainty scores. Next, we employ a physics-informed analysis to evaluate the quality of each segmentation result (Fig 1-A). Specifically, outliers are identified by analyzing the temporal behavior of myocardial pixel time-curves, by measuring deviations in the centroid of the area under the time curve (Fig 1-B). The final segmentation solution is selected based on the lowest outlier score. A total of 106 patients from 3 centers in the registry we included in the external test dataset as described in Fig 2-A. To evaluate the performance of the proposed hybrid physics-informed approach vs. the uncertainty-based SOTA technique, we focused on two common segmentation errors in FPP images that are difficult to detect with Dice score [6]: Type I error, where bloodpool is mistakenly included in the segmented myocardium; and, Type II error, where areas with minimal blood flow (e.g., epicardial fat) are erroneously included in the segmentation.
Results:
The proposed hybrid model selection approach demonstrated significantly improved performance compared to the SOTA method. The prevalence of Type I errors was reduced from 20.3% to 4.0%, and Type II errors were reduced from 7.7% to 1.0% (p < 0.001 for both). The improvement was most notable in cases where the SOTA approach resulted in subtle, yet critical segmentation errors as shown in the two representative cases presented in Fig 3, where the proposed physics-informed approach correctly excluded erroneous regions that the SOTA approach included, thereby improving the accuracy of myocardial segmentation.
Conclusion:
Our results, which leverage multi-center perfusion CMR datasets from the SCMR registry, suggest that combining physics-informed guidance and uncertainty-based DNN model selection significantly reduces the prevalence of segmentation errors in deep learning-based analysis of multi-center stress FPP datasets. This hybrid approach has the potential to improve the reliability of fully automated stress FPP analysis in clinical settings and in multi-center clinical trials.
