Automatic classification of cardiomegaly using deep convolutional neural network

Chest radiograph is a primary imaging technique to detect cardiomegaly, a condition where the heart is enlarged. A fast reliable diagnosis is crucial as it would assist the clinicians to decide on the right treatment options. In this thesis, a deep learning method is applied to train a convolutional neural network to automatically classify cardiomegaly from chest radiograph images. Furthermore, a generative adversarial network (GAN) is trained to synthesize additional chest radiograph images. The images are later used for GAN-based augmentation, by which possible enhancement of the performance of automatic classification is investigated. The pipeline is completed by utilizing the Puhti supercomputer owned by CSC IT center in Finland. By transfer learning methods, VGG16 network architecture with pre-trained weights from the ImageNet dataset is utilized to train the classification model. Evaluation of this model on a test set resulted in an AUC of 0.845 and an accuracy level of 77.1%. However, the performance of the classifier was not significantly improved by GAN-based augmentation technique. Future work and suggestions for improvements have been provided based on observations of results and data