Radiomics-based detection of acute myocardial infarction on noncontrast enhanced midventricular short-axis cine CMR images
Baptiste Vande Berg,Frederik De Keyzer,Alexandru Cernicanu,Piet Claus,Pier Giorgio Masci,Jan Bogaert,Tom Dresselaers
DOI: https://doi.org/10.1007/s10554-024-03089-9
2024-04-19
The International Journal of Cardiovascular Imaging
Abstract:Cardiac magnetic resonance cine images are primarily used to evaluate functional consequences, whereas limited information is extracted from the noncontrast pixel-wise myocardial signal intensity pattern. In this study we want to assess whether characterizing this inherent contrast pattern of noncontrast-enhanced short axis (SAX) cine images via radiomics is sufficient to distinguish subjects with acute myocardial infarction (AMI) from controls. Cine balanced steady-state free-precession images acquired at 1.5 T from 99 AMI and 49 control patients were included. First, radiomic feature extraction of the left ventricular myocardium of end-diastolic (ED) and end-systolic (ES) frames was performed based on automated (AUTO) or manually corrected (MAN) segmentations. Next, top features were selected based on optimal classification results using a support vector machine (SVM) approach. The classification performances of the four radiomics models (using AUTO or MAN segmented ED or ES images), were measured by AUC, classification accuracy (CA), F1-score, sensitivity and specificity. The most accurate model was found when combining the features RunLengthNonUniformity, ClusterShade and Median obtained from the manually segmented ES images (CA = 0.846, F1 score = 0.847). ED analysis performed worse than ES, with lower CA and F1 scores (0.769 and 0.770, respectively). Manual correction of automated contours resulted in similar model features as the automated segmentations and did not improve classification results. A radiomics analysis can capture the inherent contrast in noncontrast mid-ventricular SAX cine images to distinguishing AMI from healthy subjects. The ES radiomics model was more accurate than the ED model. Manual correction of the autosegmentation did not provide significant classification improvements.
radiology, nuclear medicine & medical imaging,cardiac & cardiovascular systems