According to the World Health Organization (WHO), nearly 65 million people worldwide are affected by epilepsy. Epilepsy is a chronic neurological disorder characterized predominantly by recurrent and unpredictable interruptions of normal brain function, called epileptic seizures. For a third of diagnosed patients seizures cannot be controlled by pharmacotherapy. For such patients, a treatment would be to perform surgical resection of the epileptogenic zone. The success of such surgeries largely depends on the accuracy of the epileptogenic zone localization. Neuroimaging, including magnetic resonance imaging (MRI) and positron emission tomography (PET), has been increasingly considered in the pre-surgical examination routine. This work attempts to enhance a computer-aided diagnosis (CAD) system for epileptogenic lesion detection, leveraging multimodal neuroimaging data, building upon the foundation laid in previous work by \cite{alaverdyan2020regularized}. The proposed system employs unsupervised deep siamese networks for learning normal brain representations from non-pathological scans, followed by a series of one-class SVM models at the voxel level to generate an anomaly score map. The model, initially tested on T1-weighted and FLAIR MRI exams, demonstrated a sensitivity of 61\%. The inclusion of PET imaging data is explored to assess its additional value. However, due to the ethical considerations surrounding radiation exposure in healthy individuals, the development and integration of synthetic PET data, simulating healthy brain scans, emerges as a potential solution. This thesis makes contributions by: (1) developing strategies for generating synthetic PET images from T1 MRI scans, demonstrating their utility in augmenting original PET data to enhance the detection capabilities of the CAD system; (2) investigating the use of synthetic PET images as substitutes for real PET data in scenarios of incomplete training sets, showing that synthetic PET significantly addresses the missing modality challenge; and (3) identifying the optimal strategy for integrating multiple imaging modalities within the detection model, with late fusion of T1, FLAIR, and PET (both real and synthetic) showing to be the most effective in terms of detection ability of epileptogenic zones. Structured into two main parts, the thesis first reviews recent deep learning advancements in medical imaging, particularly in brain image analysis, existing fusion strategies in medical imaging, methods to tackle missing data challenges including synthetic image generation, and CAD systems for brain pathology and epilepsy detection. The second part focuses on experiments for PET synthesis from T1 MRI images, the application of out-of-distribution (OOD) techniques to validate the similarity of synthetic to real images, and the detailed experimental results showcasing the improved performance of the epilepsy CAD model with the integration of synthetic PET data.