Multislice left ventricular ejection fraction prediction from cardiac MRIs without segmentation using shared SptDenNet

We propose a spatiotemporal model for cardiac magnetic resonance images (MRI) named SptDenNet. The proposed model is based on DenseNet and extracts spatial and temporal features simultaneously to exploit three-dimensional information on the heart over the cardiac loop cycle. To balance the model performance and efficiency, we construct a shared end-to-end framework, in which all frames of each selected short-axis (SAX) view slices are input to SptDenNet individually to extract spatiotemporal features. Then, the extracted features of all selected SAX view slices of a patient are concatenated and input to the subsequent fully connected layer and then a softmax layer to predict the left ventricular ejection fraction directly. To address the problem of class imbalance, we use FocalLoss function by reshaping the standard cross-entropy loss such that it down-weights the loss assigned to well-classified samples. We validate our proposed framework on the Second Annual Data Science Bowl dataset. Our prediction for the left ventricular ejection fraction obtains results comparable with state-of-the-art end-to-end approaches but without segmentation. The average mean absolute error of the ejection fraction is 6.84. To further verify the effectiveness of the proposed framework, we use 4-chamber view images from the same dataset to predict the cardiac function; we obtain an accuracy of 86.07%. Our approach constructs an end-to-end model to predict the ejection fraction automatically without using image segmentation, which helps reduce manual work. Moreover, the proposed approach is computationally efficient.