The Intelligent Combination of Automated Acquisition and Analysis (IC-AAA) Study: Comparison of fully automated acquisition and analysis vs standard human-based CMR acquisition and analysis.

Presenting Author(s)

• 

  Revi Adheriyani, MD

  Research Fellow
  Centre Hospitalier Universitaire Vaudois, CHUV
  Lausanne, Vaud, Switzerland

Primary Author(s)

• 

  Revi Adheriyani, MD

  Research Fellow
  Centre Hospitalier Universitaire Vaudois, CHUV
  Lausanne, Vaud, Switzerland

Co-Author(s)

• SH

  Simon Halstensen

  Computer Vision Engineer
  ADIS SA, Switzerland
• RE

  Raphaël Erard, PhD

  Engineer
  Adis SA, Advanced interactive systems, Lausanne, Switzerland
  Lausanne, Vaud, Switzerland
• MZ

  Meng Zhang, MD

  MD
  CHUV – Lausanne University Hospital
  Lausanne, Vaud, Switzerland
• AM

  Ambra Masi, MD

  Cardiologist
  University Hospital (CHUV) and University of Lausanne (UNIL), Switzerland
• UW

  Udaranga Wickramasinghe, PhD

  Engineer
  Adis SA, Advanced interactive systems, Lausanne, Switzerland
  Lausanne, Vaud, Switzerland
• DR

  David Rodrigues, MSc

  Deputy chief radiology technician
  Lausanne University Hospital and University of Lausanne, Switzerland
• SK

  Soudabeh Kargar, PhD

  Scientist
  Rockies MRI Solutions LLC
  Aurora, Colorado, United States
• HB

  Hussein Balan, PhD

  Engineer
  Adis SA, Advanced interactive systems, Lausanne, Switzerland
  Lausanne, Vaud, Switzerland
• PF

  Pascal Fua, PhD

  Professor
  Computer Science, Federal Institute of Technology Lausanne, EPFL, Lausanne, Switzerland
  Lausanne, Vaud, Switzerland
• 

  Juerg Schwitter, MD

  Full Professor
  University Hospital and University of Lausanne
  Lausanne, Vaud, Switzerland

Background: CMR is well-accepted for the work-up of cardiac patients and is increasingly used which requires higher scanning capacities. Two major factors limiting this development are 1). the need for highly trained CMR operators and 2). physicians specialized in CMR data analysis. Training of CMR operators and physicians is labor-intensive and costly. The use of artificial-intelligence (AI)-based algorithms for acquisition and analysis in an automated fashion is attractive. Such algorithms were developed and tested in this pilot study.

Methods: Development of scanning and analyses algorithms: First, an image acquisition protocol was defined on a remote host (Adis SA, Lausanne) which was then connected to the MR scanner (Aera, 1.5T, Siemens Healthineers) via access-i (syngo MR XA20, Siemens) which allows for MR scanner remote control and bi-directional data transfer. First, localizers were automatically acquired and sent to the remote host to automatically reconstruct a 3D heart model for left-ventricular (LV) long-axis determination. With this information the algorithm on the remote host started the short-axis (SAX) cine breath-hold (BH)-acquisition with pauses after each slice acquisition to allow the operator for BH instructions. For the automated analysis of the cine LV SAX stack, dynamic patient-specific 3D heart models were constructed by combining segmentation, contour registration, geometric deep learning, and classical registration (Fig 1A,B).¹

To compare the fully automated acquisition and analysis (AAA) technique with the current standard, 10 healthy volunteers were studied. The exam started with the automated acquisition. Then, the operator took over for the manual acquisition of localizers and a stack of BH cine SAX images. The exam ended with an aortic flow acquisition. The volunteer was then removed from the bore and newly repositioned for a second MR examination using the same protocol as in examination 1. Imaging parameters for the LV cine SAX stack were TR/TE 36.92/1.19ms, spatial resolution 1.4x1.4mm², slice thickness/gap: 8/2 mm.

The cines stacks were analyzed fully automatically by the analysis algorithm¹ and also by 2 physicians blinded to the acquisition mode to yield LV volumes and function.

Results: Time for LV long-axis determination was shorter with the fully automated mode (n=10, -2min23s±1min3s, p< 0.001). The long-axis angles were highly reproducible by both methods. The AAA-mode angle was 23°±5° steeper (to the feet) and 10°±10° more central (both p< 0.01). Fig. 2 and the Table show the good agreement between the 2 methods with small biases resulting in a better agreement of the AAA method with aortic flow (Fig. 1C).

Conclusion: AI-based fully automated acquisition and analysis of LV volumes and function in healthy subjects yield results comparable to those obtained by humans. The fully automated AAA-technique shows better agreement between LV stroke volume and aortic flow than the standard human-based technique.

Comparison of Standard Human-based Acquisition and Analysis vs AAA-Mode
Figure 1A: Fully automatically generated contours overlaid on fully automatically acquired cine short-axis loops. The only interaction with a human operator was to instruct the volunteer to hold his breath. Images represent end-diastolic phase. Figure 1B: Images represent end-systolic phase. Figure 1C: The human-based standard acquisition and analysis yields left ventricular stroke volume (LVSV) slightly overestimating aortic flow. Overestimation by the AI-based LVSV is less vs human-based LVSV (-9.1±6.7 ml; p<0.0021). Of note, the 3D-heart modeling is only trained for LV shapes so far.
Figure 2: Bland-Altman plots: The grey area corresponds to a coefficient of variation of 10% (CV%). If the 95%-confidence interval is within this grey area, the measurement fulfills the generally accepted CV% indicating its potential usefullness in clinics, i.e. a replacement of the standard method by the novel method for left ventricular end-diastolic volume (LVEDV), LV stroke volume (LVSV), and LV éjection fraction (LVEF). LVESV: LV end-systolic volume.