Publications

Abstract : Several medical therapies like tumor segmentation depend upon the fine resolution of the imaging. The severe dependency on CT images limits the advancements of the field, which can be overcome by medical imaging translation from CT to MRI images. The present work offers an analysis of how to perform cross-modality transfer between two medical modalities, computed tomography (CT) scans and magnetic resonance imaging (MRI) data using CycleGAN neural processing method. We performed this analysis on two different datasets of sizes 40 and 367, respectively. The effect of dataset size and the information it contains have been examined. It is found that the CycleGAN can learn robust features even for a smaller size dataset; however, the quality of the model improves with the dataset size for a given number of epochs. Average values of PSNR/SSIM are found to be 41.2/0.67 and 75.3/0.23 for the models developed on small and big datasets, respectively. Irrespective of the low quality of image translation, the present work is useful for medical image data augmentation, which is further helpful in improving the efficiency of other neural network-based medical tasks such as segmentation.