Cardiovascular Disease Prediction through Ensembled Transfer Learning on Cardiac Magnetic Resonance Imaging

Cardiovascular Diseases (CVD) cause more deaths worldwide than most of the other diseases. The diagnosis of cardiovascular disease from Magnetic Resonance Imaging plays a major role in the medical field. The technological revolution contributed a lot to increase the effectiveness of CVD diagnosis. Many Artificial Intelligence methods using Deep Learning models are available to assist the cardiologist in the diagnosis of CVD from Magnetic Resonance Imaging (MRI). In this study, we leverage on the merits of deep learning, transfer learning, and ensemble voting to improve the accuracy of Artificial Intelligence-based CVD detection. VGG16, MobileNetV2, and InceptionV3, trained on ImageNet, are the models used and the dataset is the Automatic Cardiac Diagnosis Challenge dataset. We customized the classification layers of all three models to suit the CVD detection problem. The results from these models are ensembled using the soft-voting and hard-voting approaches. Test accuracies obtained are 97.94% and 98.08% from hard-voting and soft-voting respectively. The experimental results demonstrated that the ensemble of outputs from transfer learning-based Deep Learning models produces much improved results for CVD diagnosis from MRI images.