Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back
On-Demand Video

ISMRM - SCMR Workshop

ISMRM - SCMR Workshop: Oral Session 5: Artificial Intelligence in Diffusion and Strain MRI

Oral Abstract 3: Comprehensive Deep Learning-Automated Analysis of Cardiac Function and Myocardial Strain Using CMR in Clinical Setting: Achieving Comparable Accuracy to Expert Observers

Thursday, January 30, 2025
8:35 AM – 8:45 AM East Coast USA Time
Location: Blue Room
Background: Recent advancements in artificial intelligence have spurred the development of automated solutions for cardiac function and strain analysis, reducing the burden of manual tasks like contouring myocardium boundaries and mitigating human error. The need for a fully automated analysis workflow is especially critical in large-scale retrospective studies and at major imaging centers, where automation can streamline process from data management to results without user input. This study investigates the effectiveness of a zero-click workflow for cardiac function and strain analysis using CMR cine images from a clinical setting. The workflow was tested with clinical data from cardio-oncology patients and validated by multiple CMR readers.

Methods: Fifty patients underwent routine CMR scans with IRB approval. The DICOM images were processed using a fully automated workflow based on deep-learning neural networks. The automation steps included cine sequence detection, view classification, short-axis (SAX) slices selection covering the left ventricle (LV), identification of end-diastole (ED) and end-systole (ES) phases, cardiac structure delineation for volumetric parameter calculation, myocardium motion tracking, and strain analysis. One senior and one junior observer independently reviewed and modified the automated results. Agreement was assessed using the Intraclass Correlation Coefficient (ICC), and comparison was performed with Bland-Altman analysis.

Results: The classification of cine sequences and imaging views showed perfect agreement between the AI workflow and observers. Minor confusions between four-chamber (4ch) and three-chamber (3ch) views had negligible effects on SAX slice selection. The average root mean square error (RMSE) for ED, ES, and SAX slice selection between observers and AI were 0.4, 0.9, and 0.8, respectively. Manual adjustments were required for only 1.9% of images. The AI analysis demonstrated high agreement with human measurements for both function and strain (Table 1). While significant differences were observed in LV EDV and LV EF between human and AI analyses, these discrepancies were minimal when relatively to the data range and did not affect clinical interpretation. Bland-Altman plots are shown in Figure 1 and example images are in Figure 2.

Conclusion: This study evaluated a fully automatic, zero-click workflow for cardiac function and strain analysis in clinical CMR patients, capable of processing raw images from an entire study without human input. All 50 studies were successfully analyzed, with the AI workflow achieving near-perfect agreement with human observers. The AI-enhanced workflow significantly reduced the reader  burden, as less than 2% of images required manual adjustments. The inter-observer variability between the AI and human observers was comparable to that observed between human observers, demonstrating consistent performance across all function and strain parameters.

Figure 1. Bland-Altman analysis of function (LV EDV, LV ESV and LV EF) and strain (GCS, GRS, GLS) measurements. Inter-observer comparisons were shown between the averaged human observers (avg Ob) and the fully-automated workflow (AI).

Figure 2. Example motion-analyzed images using the fully-automated workflow. Epicardium and endocardium boundaries were automatically detected and tracked through all cardiac phases.