Optimization of Generative Models using No-Reference Metrics: Application to MRI to CT Translation

In the field of medical imaging, the application of generative artificial intelligence techniques for cross-domain translation represents a research area of significant interest. Nevertheless, a comprehensive review of the literature has revealed a lack of established metrics that can guarantee the reliability and quality of generated images. This thesis investigates the use of the pix2pix image translation model for the transformation of medical images from Magnetic Resonance Imaging (MRI) to Computed Tomography (CT). The goal is to identify objective metrics for an accurate evaluation of the produced images. In addition to the utilisation of conventional Full-Reference metrics, including MAE, MSE, PSNR and SSIM, No-Reference metrics such as NIQE, ILNIQE and PIQE were also examined. Moreover, NIQE and ILNIQE were introduced as loss functions during the model training process, with the objective of improve the quality of the generated images. The results demonstrated the challenge of using both Full-Reference and No-Reference metrics to assess the quality of the synthesised images, although ILNIQE showed particular promise. However, the application of ILNIQE as a loss function exhibited limitations due to the high computational time required, whereas NIQE, in combination with other traditional loss functions, produced satisfactory outcomes. It can be concluded that further developments are required in order to validate and improve the reliability of the translated images.