Optimizing Network Classification Performance by Geometric Transformations on Delayed Enhancement Cardiac Magnetic Resonance Imaging

Delayed enhancement cardiac magnetic resonance imaging is crucial in identifying and monitoring heart disease. Since Deep Convolutional Neural networks have been found to perform very well in different computer-assisted activities, the use of these automated methods appears to have potential for reducing the workload of radiologists and improving workflow efficiency. Nevertheless, these networks rely significantly on big data to avoid biases and accurately learn the feature conditions. To address this issue, the use of data augmentation techniques has been suggested. In this work, we develop an automated deep-learning method to assist radiologists in classifying the left ventricle segment in cardiac MRI images by using pre-trained convolutional neural networks. Four popular network architectures, namely GoogLeNet, SqueezeNet, ResNet-50 and ShuffleNet were compared, and the abilities of these networks to perform the task were examined on augmented data using geometric transformation. All network models were trained and tested on 80% and 20% of the images, respectively, using five-fold cross-validation. On the augmented dataset and the same training network parameter, ResNet50 architecture achieves the highest performance with an average accuracy of 97.78% and F1-score of 0.9776. All networks’ performances differ slightly from one another. The finding shows that the target class, which is the LV segment, performs exceptionally well after the geometric transformation augmentation technique has been applied.