Editorial for “Multisequence <scp>MRI</scp> Enables High‐Fidelity <scp>FDG</scp> ‐ <scp>PET</scp> Synthesis for Epilepsy Using <scp>GANs</scp> ”

Deep learning has enabled multimodal medical image transformation, and, in particular, Generative Adversarial Networks (GANs) have demonstrated the ability to generate realistic synthetic images from one imaging modality to another. The earliest application is MRI-to-CT, partly due to the development of hybrid MR/PET systems that require accurate attenuation correction, which can only be obtained on CT [1]. Following its success, the approach was widely implemented for transformation and reverse transformation across CT, MRI acquired with different sequences, and PET acquired with various tracers. Essentially, when large sets of paired images are available, the Pix2Pix GAN can generate synthetic images that appear to be anatomically reasonable. The evaluation was primarily based on visual assessment of the image contrast on synthetic images, and using quantitative metrics such as the Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), and Normalized Root Mean Squared Error (NRMSE) calculated using the ground truth image as the reference. However, there are no established criteria for determining whether the quality of the synthetic images is sufficient for clinical applications. For pathological conditions such as tumors or epilepsy that require accurate localization of small lesions, the GAN approach was skeptical. Anatomically normal-appearing images may not reveal the subtle abnormalities. Properly designed, rigorous validation studies are needed to establish the role of synthetic images and their reliability for use in patient management. Several studies have attempted to generate synthetic PET images from MRI for brain tumors [2, 3] and epilepsy [4, 5] studies. However, most of these were only aiming to demonstrate feasibility using a small number of patients. In addition to the general assessment of image quality, the lesion detection capability, the feature similarity index measure (FSIM), and the lesion-to-background ratio measured on synthetic PET images can be used as evaluation metrics. In patients with epilepsy, due to temporal variations, acquiring simultaneous PET/MR images would provide the absolute ground truth PET corresponding to the brain state at the time of MRI acquisition. Such studies were challenging to perform. This article by Lu et al. [6] was well executed and introduced several notable innovations not yet reported in the literature. First of all, the input from MRI to generate PET includes three channels: the T1-weighted images and two functional maps, fractional amplitude of low-frequency fluctuation (fALFF) and regional homogeneity (ReHo), derived from resting-state functional MRI (rs-fMRI) previously shown by the authors to be associated with hypometabolism measured by FDG-PET [7]. Second, there were 481 cases (311 for training/validation and internal test, 170 for external test), and each patient underwent MRI and FDG-PET acquisition simultaneously on a hybrid PET/MR system. This is the largest MRI-PET GAN dataset reported to date, with real PET as the ground truth for all subjects. Third, the external set included 115 patients with temporal lobe epilepsy (TLE) who underwent surgery, and the outcomes were available by determining them from a one-year follow-up. The great dataset enabled the design of rigorous validation studies. The authors first demonstrated that the synthetic PET had Excellent/Good image quality based on visual assessment and quantitative metrics (SSIM, PSNR, MSE, and NRMSE), and that the measured standardized uptake value ratio (SUVR) across brain regions, when referenced to the cerebellum, was not significantly different between real PET and synthetic PET. These were the evaluation methods commonly reported in the literature. Subsequently, they calculated an asymmetry index and showed that both real PET and synthetic PET demonstrated significant differences between TLE and non-TLE patients. Next, they performed a clinical assessment of the temporal lobe hypometabolic region by three readers using synthetic PET, achieving 90% accuracy in the internal test and 87% accuracy in the external test compared with ground truth PET results. Lastly, in surgical patients, the radiomics models derived from real PET and synthetic PET showed comparable classification performance in differentiating hippocampal sclerosis and predicting the postoperative outcome of Engel IA vs. IB-IV. These validation results demonstrate the potential of synthetic PET to determine the TLE lobe, identify hypometabolic regions, diagnose hippocampal sclerosis, and predict surgical outcomes. Although the dataset was impressive and the obtained results were promising, the synthetic PET was still generated from MRI information rather than from real metabolic data. The authors previously showed that fALFF and ReHo maps derived from rs-fMRI were associated with SURV; however, it remains challenging to use rs-fMRI maps for diagnosis. Incorporating them with T1-weighted images to generate synthetic PET images effectively consolidated all information into a PET image format familiar to nuclear medicine physicians. The authors emphasized that synthetic PET cannot replace the acquired PET. This is the consensus of the medical imaging community that artificial intelligence algorithms can be applied to improve the quality of subtle contrast, but they cannot reliably generate pathology-related contrast. The presented cases that showed a small lesion clearly elaborated the limitations. In a preprint by Lucas et al. [5], the authors sought to integrate perfusion measurements obtained with arterial spin labeling with anatomic MRI to generate synthetic PET images for epilepsy management. It will be interesting to read their forthcoming study and to see how the validation will be performed. Another common application of GANs is to generate contrast-enhanced MR images based on non-contrast images to avoid the administration of gadolinium-based contrast agents [8]. Although published studies report that realistic synthetic post-contrast images can be generated, no rigorous validations have been conducted, and the medical imaging community remains skeptical of this approach. Artificial intelligence algorithms are capable of enhancing contrast, but can't generate contrast from limited training. Even in FDA-cleared software such as AiMIFY, which is approved for selectively amplifying brain lesion MR image contrast, strong safety caution is advised: the standard post-contrast image must always be reviewed first, and the AI-enhanced images should not be used alone to assist patient diagnosis. In summary, although GANs have been shown to generate image-to-image transformations across CT, MRI, and PET with reasonable synthetic image quality, carefully conducted clinical validation studies are needed to evaluate their reliability, determine their roles, and assess how they can assist in patient management.