Exploring Classification of SPECT MPI images applying convolutional neural networks

The main goal of this research paper is to address the problem of SPECT myocardial perfusion imaging (MPI) diagnosis, exploring the capabilities of convolutional neural networks (CNN). Up to date, very few research studies have been conducted regarding the application of machine learning algorithms focusing on efficient structures of convolutional neural networks (CNNs) for the diagnosis of ischemia in MPI images. In the presented work, the dataset consists of SPECT images in stress and rest representation, and a two-class classification problem corresponding to 262 normal and 251 ischemic cases is explored. The data augmentation technique was used for increasing the number of the training dataset by rotating the images and zooming randomly. In this research study, a simple but robust CNN model for automatic classification of MPI images in two categories, was applied, after a proper exploration process concerning different values for number of layers, dense nodes, convolutional parameters as well as batch size and pixel size. The proposed CNN achieved an accuracy of 90.2075% and an AUC value of 93.77%. The results proved that the convolutional neural network is able to differentiate between normal and ischemic cases and would be a great assist to medical industry, when researching myocardial perfusion images. The model we are proposing is considered a valuable asset for this medical classification problem, as it manages to produce more reliable results compared to traditional clinical methods.