A Patch-Wise Generative Adversarial Network for PET-MR Image Generation with Feature Attribution for Detection of Focal Cortical Dysplasia

Tese de mestrado integrado, Engenharia Biomédica e Biofísica (Engenharia Clínica e Instrumentação Médica) Universidade de Lisboa, Faculdade de Ciências, 2022More than 50 million people worldwide suffer from epilepsy with a third of those being diagnosed with drug-resistant epilepsy where the seizures cannot be treated through pharmacotherapy. In these cases, surgical removal of the epileptic brain tissue in patients is presented as an effective solution for treatment. However, for surgery success, it is vital that the accurate location of epileptic regions in the brain are known. Neuroimaging, specifically magnetic resonance imaging (MRI) and positron emission tomography (PET), commonly are the doctor’s allies in identifying these lesions’ locations responsible for the seizures. Focal cortical dysplasias (FCDs) are the most common type of cortical lesions respon sible for drug-resistant epilepsy in children. These lesions have highly heterogeneous masses, occur in different brain regions and result in different levels of visibility, corresponding to the second most in tractable type of lesion in adults with epilepsy. Moreover, among drug-resistant epilepsy cases, a third of these lesions cannot be correctly identified by neuroimaging experts, resulting in unsuccessful surgical planning and consequently ineffective treatment for patients. Recently, Generative Adversarial Networks (GANs) have demonstrated their value in neuroimaging anomaly detection. Therefore, this work pro poses the application of two different GAN methods – WGAN and CycleGAN - for anomaly detection of FCDs, in PET-MRI data of epileptic patients. A 3D patch-basis anomaly detection approach was therefore developed, inspired by previous works, to detect FCDs location by deconfounding acquisition noise and normal cortical variabilities in PET-MR brain scans of epilepsy patients. Therefore, the GAN models applied two different approaches for lesion detection: detection through reconstruction (WGAN) and detection through translation (CycleGAN). Moreover, the combination of PET and MR modalities was studied and compared to training the networks with individual imaging modalities instead. Through the results, it was possible to understand and correct some issues GAN models have when training with multimodal 3D data. However, both methods for anomaly detection were able to detect diseased brain areas in patients with very visible FCDs, although failing to identify them in patients with very subtle lesions. Recent studies will be briefly discussed in the conclusion, which propose new approaches and architectures for multimodality training, with great potential to improve the performance of the networks for anomaly detection in future works