1353465 - Unsupervised Machine Learning of Stress Perfusion CMR on a Voxel-wise Basis Correlated With Clinical Diagnosis of Myocardial Ischemia

Background: Since first-pass perfusion imaging requires temporally repeated measurements, the resulting time course data differs in interpretive style from other clinical imaging. Many interpreting imagers will be less familiar with viewing repeated measurements showing spatially dependent temporal shifts in signal intensity that would indicate a first-pass hypoperfusion defect consistent with ischemia.  Unsupervised machine learning (UML) makes no assumptions about the ‘expected’ behavior of signal; patterns shared between observations, such as different voxels, are clustered mathematically. UML differs from supervised deep learning in that no training is required, and the results are based only on a multivariable linear relationship between input and output. The purpose of this work is to investigate whether the myocardial perfusion imaging data processed by UML is correlated with the clinical findings of myocardial ischemia and how differences in input data affect the algorithm’s result.

Methods:

Six patients with the diagnosis of myocardial ischemia based on adenosine stress cardiac MRI were included and six age-matched controls using the same scan parameters without evidence of ischemia or scar/infiltrative disease.  Perfusion images of the entire acquired field-of-view (FOV_a) and from a segmented myocardial region-of-interest (ROI_m) were analyzed with UML [Figure 1] on a per-voxel basis. First, the data were dimensionally reduced by principal component analysis, recording the number of components that captured at least 1% of the overall observed variance. Second, clustering was performed using a Gaussian mixture model with optimum number of clusters determined by Calinski-Harabasz criteria. The differences in dimensionality reduction characteristics and differences in optimal number of clusters between groups based on input type and based on ischemic status were compared by t-test.
 


Results:

Dimensionality reduction was performed similarly on all 12 patients (average age = 72) for the FOV_a data input, regardless of the diagnosis of ischemia, with 6.1  +/- 5.4 dimensions required to reach 99% variance-explained threshold. Dimensionality reduction also performed similarly within the same 10 patients for the ROI_m input, requiring 2.6 +/- 1.7 dimensions to capture 99% of variance. Significantly fewer dimensions were required for ROI_m than for FOV_a (p=0.01). Dimensionally reduced FOV_a images also demonstrated substantial visual difference among the first 3 principal components, mapped by color [Figure 2].

Cluster analysis by visual analysis separation be ischemic and non ischemic myocardium.

Conclusion:

From entire acquired field-of-view stress perfusion MR inputs, UML showed the ability to cluster by tissue type, while segmented myocardial field-of-view inputs were able to cluster ischemic and non-ischemic myocardium. This suggests a possible role for UML in the analysis pipeline for perfusion CMR, particularly in an era of increasing automatic segmentation of myocardium.