Deep Learning-Based Cardiac MRI Planning from Localizers to Cine Views Using Landmark Detection.
Authors
Affiliations (5)
Affiliations (5)
- Department of Electrical and Computer Engineering, University of Iowa, Iowa City, Iowa (D.D.D., A.R.).
- Department of Radiology, University of Iowa Carver College of Medicine, Iowa (S.G., O.F.D.P., T.R., S.P.).
- Department of Radiology, University of Miami, Miller School of Medicine, Miami, Florida (P.Y.A.).
- Department of Radiology, University of Wisconsin School of Medicine and Public Health, Madison, Wisconsin (P.N., S.P.).
- Department of Radiology, University of Iowa Carver College of Medicine, Iowa (S.G., O.F.D.P., T.R., S.P.); Department of Radiology, University of Wisconsin School of Medicine and Public Health, Madison, Wisconsin (P.N., S.P.). Electronic address: [email protected].
Abstract
This study evaluates a fully automated deep learning framework to enhance the efficiency and accuracy of cardiac MRI planning. In this retrospective study, data from 1023 patients (ages 8-90 years) who underwent cardiac MRI were analyzed, including coronal, sagittal, axial localizers, and short-axis (SAX) and long-axis cine images. Experts manually annotated landmarks, serving as the ground truth for developing deep learning models. The models were assessed using 5-fold cross-validation. Performance metrics included median landmark distances and plane angle differences. The model achieved robust performance in landmark localization across all cardiac MRI planes. For localizer images, median distances were 5.1 mm (superior) and 7.2 mm (inferior) on coronal views, and 5.6 mm (superior) and 7.5 mm (inferior) on sagittal views. Median distances for axial, 2-chamber, and 4-chamber landmarks were 5.2 mm, 5.2 mm, and 5.6 mm, respectively. In short-axis mid slices, annotations based on the left ventricular center, right ventricular insertion points, and right ventricle obtuse angle had a median error of 5.2 mm, while basal slice valve-based annotations had 4.6 mm error. Angular deviations for SAX planning were 2.0° (2CH) and 1.5° (4CH). For long-axis views, angulation errors were lower using SAX mid slices (3.3° for 2CH, 2.6° for 4CH) compared to SAX base (4.0° and 3.9°, respectively). A deep learning-based automated workflow for cardiac MRI planning is feasible with improved precision.