GAN-Based Synthetic FDG PET Images from T1 Brain MRI Can Serve to Improve Performance of Deep Unsupervised Anomaly Detection Models
Résumé
Research in cross-modal translation or synthesis domain has been very productive over the past few years to tackle the scarce availability of large curated datasets for the training of deep models, with promising performance of GAN-based architectures. However, only a few of these studies assessed task-based related performance of these synthetic data. In this work, we design and compare different GAN-based frameworks for generating synthetic brain FDG-PET images from T1weighted MRI data, and explore further impact of adding these fake PET data in the training of a deep brain anomaly detection model. Qualitative and quantitative results allow us to conclude that the generated PET images look similar to real ones with SSIM and PSNR values around 0.88 and 23.5 respectively for the best GAN architecture. Training of the brain anomaly detection model on hybrid datasets including 35 real and 40 synthetic FDG PET data, allows achieving a 65% detection sensitivity of subtle epilepsy lesions in 17 real PET exams of patients, while the sensitivity is 53% when training with the 35 real PET exams only, thus demonstrating the diagnostic value of these synthetic data for the design of CAD models.
Origine : Fichiers produits par l'(les) auteur(s)