Imaging Epilepsy: Past,Passing,and to Come

Abstract

New imaging techniques appearing over the last few decades have replaced procedures that were uncomfortable, of low specificity, and prone to adverse events. While computed tomography remains useful for imaging patients with seizures in acute settings, structural magnetic resonance imaging (MRI) has become the most important imaging modality for epilepsy evaluation, with adjunctive functional imaging also increasingly well established in presurgical evaluation, including positron emission tomography (PET), single photon ictal-interictal subtraction computed tomography co-registered to MRI and functional MRI for preoperative cognitive mapping. Neuroimaging in inherited metabolic epilepsies is integral to diagnosis, monitoring, and assessment of treatment response. Neurotransmitter receptor PET and magnetic resonance spectroscopy can help delineate the pathophysiology of these disorders. Machine learning and artificial intelligence analyses based on large MRI datasets composed of healthy volunteers and people with epilepsy have been initiated to detect lesions that are not found visually, particularly focal cortical dysplasia. These methods, not yet approved for patient care, depend on careful clinical correlation and training sets that fully sample broad populations.

Introduction

In the era of pneumoencephalography (PEG) and angiography, imaging gave neurologists caring for people with epilepsy only limited, and often nonspecific information, with considerable discomfort and some risk.^1,2 For example, PEG in 45 children with epilepsy revealed 33 normal studies; seven showed “slight cerebral atrophy,” one a porencephalic cyst, and four were “unsuccessful.” There were 13 complications, including the child with the cyst, who developed persistent hemiparesis and cognitive impairment. ¹ Even magnetic resonance imaging (MRI) seemed disappointing initially before appropriate sequences were developed for identification of lesions such as mesial temporal sclerosis. ³ Now we have reached an exciting, but also fraught era. Technical advances have led to striking anatomic and functional visualization, but the increasing sophistication of analysis models may put clinical neurologists at a disadvantage when attempting to understand and interpret what they see. Some studies, such as positron emission tomography (PET) neurotransmitter receptor imaging, remain more appropriate in research than standard clinical settings. ⁴ Moreover, the plethora of available scans can lead to anxiety over which to choose, and how to align the results with data from other modalities such as electroencephalography (EEG).

Clinical Epilepsy Imaging in 2025

Neuroimaging plays a critical role in clinical evaluation, diagnosis, and management of patients with epilepsy. Immediate noncontrast head computed tomography (CT) is still often the study of choice in the acute setting for evaluation of new-onset seizures due to its rapid and widespread availability. Head CT is of higher yield in patients presenting with an abnormal neurologic examination, predisposing history, and/or focal seizure onset, allowing for identification of abnormalities such as tumors, hemorrhages, or infections, and leading to changes in management in 9–17% of adult and 3–8% of pediatric patients. ⁵ When available, however, structural brain MRI is the imaging modality of choice as it is more sensitive for identification of epileptogenic lesions and does not involve exposure to radiation. Imaging can be obtained nonurgently in patients who are not acutely ill and have returned to baseline. Overall yield is low in children over two years of age with generalized seizures, normal examination, and EEG normal or with generalized abnormalities, and may not justify the anesthesia risk, discomfort, and expense. Similarly, the yield of neuroimaging in patients presenting with recurrent seizures is low in the absence of acute head trauma, prolonged alteration of consciousness, or focal neurologic examination.

Optimized protocols for imaging patients with epilepsy have been recommended by the International League Against Epilepsy, including at a minimum isotropic millimetric 3D T1-weighted and fluid-attenuated inversion recovery (FLAIR) images and high-resolution 2D sub-millimetric T2 images. ⁶ Gadolinium-enhanced and susceptibility-weighted imaging are suggested if there is suspicion of tumor, vascular malformations, infection, or neurocutaneous syndromes. Identifying epileptogenic lesions with structural MRI following a first seizure can support a diagnosis of epilepsy, inform decision-making regarding treatment with antiseizure medications, and inform prognosis. ⁷

Structural and functional imaging play a critical role in evaluating patients with drug-resistant epilepsy for surgical interventions. Epileptogenic lesion identification is associated with significantly improved postoperative seizure outcomes following surgical resection. ⁸ Optimized imaging protocols, including additional sequences such as double inversion recovery and fluid and white matter suppression may improve identification of subtle epileptogenic lesions, particularly focal cortical dysplasias (FCDs). ⁹ Increasing MRI field strength from 1.5 T to 3 T to 7 T is associated with increased yield, ¹⁰ as is adoption of image postprocessing machine learning approaches utilizing voxel or surface-based analyses. ⁶ A 7 T MRI provides better imaging contrast and spatial resolution, presenting technical challenges including greater inhomogeneity and artifacts, longer scans with greater subject claustrophobia due to a smaller head coil and longer bore, peripheral nerve stimulation, dizziness, and uncertain implant safety. ¹¹

Adjunctive functional imaging approaches continue to be important for presurgical evaluation, particularly in patients without clear epileptogenic lesions and/or discordant or inconclusive MRI and EEG findings. Functional abnormalities tend to be more regional than structural abnormalities and are state-dependent, with results often dependent on timing in relation to the most recent seizure. The most widely used adjunctive imaging approaches are identification of hypometabolism using fluorodeoxyglucose PET (FDG-PET) and relative hyperperfusion using single photon ictal–interictal subtraction computed tomography coregistered to MRI (SISCOM). PET and SISCOM alterations can be identified in 79%–95% of patients with mesial temporal lobe epilepsy (TLE), with decreased sensitivity in extratemporal epilepsy. ¹² Additional less widely used modalities include identification of perfusion changes using arterial spin labeling MRI and metabolism using magnetic resonance spectroscopy. EEG, magnetoencephalography source imaging, and simultaneous EEG-functional MRI (EEG-fMRI) can be used to localize sources of epileptiform activity and aid presurgical focus localization. ¹³ However, these techniques depend on assumptions embedded within modeling approaches and require significant analysis expertise. Task-based fMRI is now widely used for functional mapping, and diffusion tensor imaging can aid with identification of critical white matter tracts to assist with minimization of postoperative neurologic deficits ¹³ (Figure 1).

Figure 1.

Imaging Drug-Resistant Epilepsy: Presurgical Evaluation.

Finding the Lesion you Cannot see

Structural epilepsy lesions are an important cause of drug-resistant focal epilepsy. Despite improvements in MRI, a proportion of lesions are subtle and overlooked on visual inspection. This proportion varies according to MRI quality, underlying histopathology, expertise of the reviewing radiologist as well as what additional information the radiologist has available. There has been considerable research aimed at automating the detection of structural lesions and in particular the “MRI negative” cases as presurgical MRI identification and complete resection of a structural brain lesion are predictors of postsurgical seizure freedom. ¹⁴

Analysis of resected specimens from patients deemed “MRI-negative” during their presurgical evaluation indicates that in adult and pediatric cohorts, FCDs account for around 45%, and hippocampal sclerosis (HS) accounts for around 10% of MRI-negative cases.^15,16

There are a number of multicenter validated, available, machine-learning algorithms for the detection of FCDs. Input features used to detect these subtle lesions vary from morphometric junction, extension, and thickness maps computed using the morphometric analysis program, ¹⁷ to surface-based morphological and intensity features, ¹⁸ and to patches extracted from T1 and FLAIR MRI data. ¹⁹ However, a common challenge for these algorithms is the identification of additional areas (false positives [FPs]), as well as the FCD, ²⁰ leading to these algorithms having low positive predictive values (PPVs). A novel approach from the Multi-centre Epilepsy Lesion Detection (MELD) project incorporates whole-brain-context using a graph-convolutional neural network to reduce the FPs, ²¹ resulting in an algorithm with a PPV of around 70%. If the algorithm detects a lesion, the likelihood that it is a true FCD is around 70%. The algorithm is able to identify 64% of “MRI-negative” lesions. This openly available tool (https://github.com/MELDProject/meld_graph) outputs an interpretable patient report detailing any identified putative lesions, a characterization of the features driving the classifier's decision, as well as a classifier confidence score (Figure 2).

Figure 2.

Overview of MELD Graph and AID-HS Machine-Learning Algorithms for “Finding the Lesions You Cannot See.”

For the automated diagnosis of HS, automated and interpretable detection of HS, (AID-HS) is an open-source tool (https://github.com/MELDProject/AID-HS) designed to detect and lateralize HS. ²² It uses a logistic regression classifier trained on the asymmetries of hippocampal morphological features extracted from T1 MRI data. On a validation cohort of 275 patients, AID-HS accurately lateralized 96%, including 92% of patients deemed “MRI-negative.”

Machine-learning algorithms have the potential to transform the care of patients with “lesions we cannot see.” Through identification of a structural lesion, patients can become good surgical candidates with a high likelihood of seizure freedom postoperatively. However, to date, no MRI-based artificial intelligence (AI) algorithms for epilepsy diagnosis have obtained regulatory approval, and tools are therefore currently used as research tools. Furthermore, when a patient with focal epilepsy has an “MRI-negative” scan, the underlying structural etiology is unknown. Ideally, automated approaches would detect multiple focal epilepsy pathologies rather than requiring individual algorithms per pathology. Future work will involve developing AI models to support the diagnosis and prognosis of multiple epilepsy pathologies.

Neuroimaging in Genetic–Metabolic Epilepsies

Neuroimaging in inherited metabolic epilepsies is integral to diagnosis, monitoring, and assessment of treatment response. Imaging follows geographical patterns that fit the localization of the disorders. ²³ Abnormalities are categorized as deep gray matter predominant (eg Leigh syndrome), cortical gray predominant (eg neuronal ceroid lipofuscinoses), white matter predominant (eg Alexander and Canavan disease), combined gray–white (eg generalized gangliosidoses), and negative on conventional imaging (eg glucose transporter-1 and creatine deficiencies) (Table 1).

Table 1.

Imaging Patterns in Neurometabolic Disease.

I. Deep gray matter predominant

Examples: Leigh syndrome, thiamine transporter-2 deficiency, glutaric aciduria-1, propionic and isovaleric acidemias, PDHC deficiency, SSADH deficiency

II. Cortical gray matter predominant

Examples: NCL, POLG disorders

III. White matter predominant

Examples: Alexander and Canavan diseases, galactosemia

IV. Combined gray–white matter

Examples: GM1/GM2 gangliosidoses

V. Negative on conventional imaging

Examples: GLUT1 deficiency, creatine deficiencies

Abbreviations: GLUT1, glucose transporter 1; GM, monosialotetrahexosylganglioside; NCL, neuronal ceroid lipofuscinosis; PDHC, pyruvate dehydrogenase complex; POLG, polymerase gamma; SSADH, succinic semialdehyde dehydrogenase.

Standard modalities are ultrasound, CT, and MRI/MR spectroscopy. ²⁴ Sonography identifies nonspecific ventricular dilation, germinolytic cysts, echogenic branching basal ganglia vessels, and echogenic white matter. CT detects calcifications. MRI is the dominant modality. Myelin is best evaluated with T1-weighted sequences in the first year compared to T2 and FLAIR later. Spin Echo T2 is sensitive to subtle cortical malformations. Diffusion-weighted imaging/magnetization transfer is sensitive for acute injury, manifesting as reduced diffusivity, versus hypermyelination/vacuolization/degeneration with increased diffusivity.

Clinical disorders are conveniently classified as vitamin dependencies, transportopathies, amino/organic acidopathies, mitochondria, urea cycle, purine/pyrimidine, lysosomal storage, neurotransmitter, large molecule, creatine deficiencies, and disordered copper metabolism and glucose homeostasis. ²⁵

Imaging findings are nonspecific but typical associations exist. Examples include ventriculomegaly, white matter atrophy, thin corpus callosum, and mega cisterna magna in pyridoxine-dependent epilepsy; basal ganglia lesions (caudate/putamen and sparing pallidum) followed by necrosis in thiamine transporter-2 deficiency; gyriform swelling and elevated lactate in mitochondrial encephalopathies; delayed myelination and restricted corticospinal track diffusion in glycine encephalopathy; diffuse cytotoxic edema with restricted diffusion and reduced apparent diffusion coefficient with elevated lactate, branched chain amino acid, and beta-keto acid peaks in maple syrup urine disease; encephaloclastic changes preventable with newly available cyclic pyranopterin monophosphate therapy in molybdenum cofactor deficiency; combined cerebellar hypoplasia and atrophy in congenital disorders of glycosylation. It is important to recognize that iatrogenic signals can also occur, such as vigabatrin-induced T2/FLAIR hyperintensities of basal ganglia, cerebellum, brainstem, and thalamus with restricted diffusion in addition to white matter spongiosis.^26,27

New directions include near-infrared spectroscopy for cerebral energetics, with preliminary data showing enhanced peak hemoglobin responses in creatine-deficient subjects, with improvement during creatine supplementation. ²⁸ In succinic semialdehyde dehydrogenase deficiency (SSADHD), there is a pallido-dentato-luysian pattern of T2/FLAIR signal. We have shown a correlation between enlarged perivascular spaces, sleep disturbances, and gamma amino butyric acid (GABA) levels implicating dysfunctional glymphatics in SSADHD. ²⁹ Magnetic resonance spectroscopy editing-pulse sequence demonstrated increased GABA/N-acetyl aspartate ratio in all regions studied. ³⁰ Cerebral PET utilizing the ligand flumazenil demonstrates overuse-dependent GABA(A) receptor downregulation that correlates with clinical changes over time, including worsening seizures.^31,32 An investigational application for assessing myelination and neuronal conduction uses g-ratio mapping (Figure 3). ³³ In this technique, determinants of myelination are estimated by quantification of the ratio between the caliber of myelinated axons to the total axonal cross-sectional size (including its encasing myelin sheath). Consistent with the murine SSADHD model, ³⁴ preliminary data demonstrate decreased myelin integrity.

Figure 3.

G-Ratio Image Map in SSADHD Subject (Left) and Age-Matched Healthy Control (Right). The Higher G-Ratio (Denser Red-Coloring) in SSADHD Corresponds to Myelin Thinning.

A remarkable update in neurometabolic therapy is the advent of enzyme replacement, gene editing, and gene replacement therapy. ³⁵ In aromatic amino acid decarboxylase deficiency, ¹⁸F-fluorodopa PET demonstrates dramatic and sustained gene uptake and dopamine synthesis in basal ganglia.

AI in Neuroimaging

The field of biomedical imaging is rapidly absorbing new AI tools with the promise of improved diagnosis, treatment, and prediction of clinical outcomes for complex neurological disorders, including epilepsy. The development of generalizable AI-based imaging tools for epilepsy needs a concurrent development of large and diverse datasets. The field of epilepsy is collectively rising to meet this development need, through multiple collaborative efforts, including the IDEAS project, Australian Epilepsy Project, Multi-center Lesion Detection project, and ENIGMA-Epilepsy. Several fundamental concepts of AI are important for epilepsy imaging (Figure 4).

Figure 4.

Global and Individual TLE Signature Derived from Latent Space. The Top Left Panel Displays the “Global” Signal of Epilepsy (SEE-GAAN) Derived from Latent Space (Blue Areas Indicate Greater Signal and Green Areas Indicate Less Signal). This Epilepsy Signal is Centered on the Limbic Circuit, with Subcortical Structures such as the Hippocampus and Thalamus, Along with the Lateral Temporal Neocortex and the Insula. The Top Right Panel Displays 3 Example Subjects Demonstrating the Epilepsy Signal at a Single-Subject Level (ie, the “Local” Signal) that have Patterns Seemingly Related to their Specific Epilepsy Characteristics, in this Case, the Side of Seizure Onset.

AI and Big Data are Synergistic

The field of AI has developed increasingly complex and often unsupervised models due to the reciprocity between AI and big data. In brief, big data leverages AI models to assist with classifications and predictions. In turn, sophisticated AI models require a massive scale of data that is diverse and representative to improve their decision-making processes. Recent data have demonstrated sizable increases in the accuracy of epilepsy diagnosis as the data in the training set increases from several hundred to over 1000 MRI scans. This increase in accuracy is particularly robust as one uses increasingly sophisticated methods (eg moving from 2D slices to 3D volumes) for training models. ³⁴

Generalizability of AI Depends on Diversity in the Training Data

A clear use case for AI-aided medical imaging is in resource-limited countries with limited medical personnel. However, most epilepsy neuroimaging datasets emerge from well-resourced countries. This lack of diversity in study samples presents a challenge for researchers seeking to develop AI-based tools that can generalize to the broader epilepsy community and specifically to some of our most vulnerable patients. Until we can expand our training data to include patients with ample sociodemographic, clinical, and neuropathological diversity, caution must be taken when applying new AI tools to groups who are not well represented in the data. Such models may be biased and lead to inaccurate classifications or predictions.

Clinical Translation in Epilepsy Depends on Transparency and Interpretability

Developing an accurate AI tool is essential, but not sufficient for clinical translation. AI tools must also be transparent and interpretable if they are to be used as clinical decision-support tools. The field of Explainable AI is burgeoning and enabling the deployment of new visualization techniques that can help clinicians interpret not only what AI tools are doing, but why they are making a decision. One example, Semantic Exploration and Explainability using a Generative Adversarial Autoencoder Network (SEE-GAAN), is a generative AI approach that derives signatures of single-patient features in a visually explainable way (Figure 4). This approach has been successfully applied to chest radiographs and is just now being applied to brain imaging in epilepsy.^35–37

The development of AI tools applied to epilepsy is gaining momentum, with efforts underway across multiple epilepsy imaging consortia. These efforts are promising, but will require continued commitments from clinicians, neuroimagers, computer scientists, and funding agencies to develop tools that are scalable, generalizable, and interpretable for effective clinical deployment.

Conclusions

It is clear that imaging methods are essential for the evaluation of seizure disorders, and there are good guidelines for clinical use. However, obstacles can impede optimal neuroimaging use. Studies should be planned carefully, with a clear understanding of how each test will help to answer clinical questions, attain accurate data, and avoid confusing results that can mislead toward diagnostic byways and dead ends. Sometimes there can be too many tests, or tests repeated too often. Some of the approaches described in this review are still being evaluated and are not yet ready for standard clinical use. They require an understanding of the limitations and pitfalls of the postprocessing methods used to interpret them. Moreover, as tests move from research to standard clinical contexts, interpretation may be simplified, and moved into radiology departments, as, for example, with FDG-PET hypometabolism estimated visually rather than measured quantitatively, with possible loss of diagnostic accuracy.

Careful use of imaging resources is important as well to support society's ability to provide care to all people with epilepsy, addressing socioeconomic and other disparities that have limited access for many. Appropriate use may help to provide targeted but increasingly expensive therapies for rare diseases identified increasingly as causes of epilepsy.