125th Anniversary: Review Article

Eur Neurol 2022;85:333–341 Received: March 9, 2022

Accepted: May 25, 2022
DOI: 10.1159/000525262 Published online: June 15, 2022

The Role of MRI in the Treatment of

Drug-Resistant Focal Epilepsy
Andrea Bernasconi Neda Bernasconi
Neuroimaging of Epilepsy Laboratory [NOEL] and Department of Neurology and Neurosurgery, Montreal

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

Neurological Institute, McGill University, Montreal, QC, Canada

Keywords ibility and provide increasingly precise predictors of clinical

Epilepsy · Magnetic resonance imaging · Hippocampal outcomes. Current efforts aiming at strengthening the com-
sclerosis · Focal cortical dysplasia · Machine learning petences of epileptologists in neuroimaging will ultimately
reduce the need for invasive diagnostics.
© 2022 The Author(s).
Abstract Published by S. Karger AG, Basel

Background: Epilepsy is a prevalent chronic condition af-

fecting about 50 million people worldwide. A third of pa- Introduction
tients with focal epilepsy suffer from seizures unresponsive
to medication. Uncontrolled seizures damage the brain, are Epilepsy is a prevalent chronic condition affecting
associated with cognitive decline, and have negative impact about 50 million people worldwide. Seizures are gener-
on well-being. For these patients, the surgical resection of ally defined as transient symptoms and signs due to ex-
the brain region that gives rise to seizures is the most effec- cessive neuronal activity; based on these manifestations,
tive treatment. Summary: Magnetic resonance imaging they can be classified as focal or generalized. Various eti-
(MRI) plays a central role in detecting epileptogenic brain le- ologies have been associated with epilepsy, including
sions. In this review, we critically discuss advances in neuro- structural, genetic, infectious, metabolic, and immune.
imaging acquisition, analytical post-acquisition techniques, Frequent structural pathologies underlying focal epilepsy
and machine leaning methods for the detection of epilepto- include traumatic brain injury, tumors, vascular malfor-
genic lesions, prediction of clinical outcomes, and identifica- mations, stroke, and developmental disorders. In a third
tion of disease subtypes. Key Message: MRI is a mandatory of patients with epilepsy, antiseizure medication is inef-
investigation for diagnosis and treatment of epilepsy, par- fective [1]. Uncontrolled seizures damage the brain [2]
ticularly when surgery is being considered. Continuous and are associated with high risks for socioeconomic dif-
progress in imaging techniques, combined with machine ficulties, cognitive decline, and mortality [3]. This review
learning, will continue to push the boundaries of lesion vis- addresses drug-resistant focal epilepsy, specifically com-

Karger@karger.com © 2022 The Author(s). Correspondence to:

www.karger.com/ene Published by S. Karger AG, Basel Andrea Bernasconi, andrea.bernasconi @ mcgill.ca
This is an Open Access article licensed under the Creative Commons
Attribution-NonCommercial-4.0 International License (CC BY-NC)
(http://www.karger.com/Services/OpenAccessLicense), applicable to
the online version of the article only. Usage and distribution for com-
mercial purposes requires written permission.
mon syndromes secondary to focal cortical dysplasia gate tissue microstructural properties and function.
(FCD), a malformation of cortical development, and tem- Among them, diffusion-weighted MRI and its analytical
poral lobe epilepsy (TLE) due to mesiotemporal lobe scle- extension diffusion tensor imaging (DTI) have been wide-
rosis, a histopathological entity that combines neuronal ly used to image the white matter [9]. Advanced models
loss and gliosis in the hippocampus and adjacent cortices. describing diffusion within distinct microstructural con-
To date, the most effective treatment of these pathologies stituents, such as high angular resolution diffusion imag-
has been resective surgery. ing [10], have furthered the ability to resolve crossing fi-
Since the inception of magnetic resonance imaging bers [11]. In addition to ease the interpretation of conven-
(MRI) in the 80s, constantly evolving hardware and se- tional diffusion metrics, such models offer more sensitive
quences have provided increasingly detailed appraisal of markers of the microstructural environment of epilepto-
brain structure and function, making it the main investi- genic lesions [12]. Other techniques such as diffusion kur-
gative platform in neuroscience of health and disease. tosis imaging have the ability to quantify intra-voxel tissue
While the investigation of epilepsy is multidisciplinary, properties [13], which may ultimately refine lesion detec-
MRI has been particularly instrumental in the presurgical tion. Neurite orientation dispersion and density imaging,
evaluation as it can reliably detect epileptogenic lesions. commonly referred to as NODDI [14], is a reconstruction

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

Indeed, localizing a structural lesion on MRI strongly technique based on a multi-shell acquisition protocol that
predicts a favorable outcome after surgery [4–6]. Yet, estimates intra- and extracellular volume fractions of neu-
challenges remain. Many patients have subtle lesions that rites (i.e., dendrites and axons). In TLE, these techniques
are undetected on routine MRI but found on histology. have unveiled gray and white matter abnormalities in re-
These patients, oftentimes referred to as MRI-negative, gions not detected by conventional DTI [15]. Other quan-
represent an utmost clinical challenge. Indeed, notwith- titative contrasts reflect actual quantities biophysically
standing long and costly hospitalizations for EEG moni- linked to tissue microstructure. T1 mapping along the cor-
toring with intracerebral electrodes (SEEG), surgery is tical mantle and hippocampal subfields has revealed al-
less likely to be performed [4]; when operated, these pa- tered myelin content in ipsilateral temporal and frontal
tients consistently show worse seizure control compared limbic regions, independent from morphology and inten-
to those with visible lesions [5]. These shortcomings have sity [16]. Interestingly, multiparametric imaging combin-
motivated the development of advanced analytic tech- ing anatomical, functional, and metabolic data can be ob-
niques for the discovery of diagnostic and prognostic bio- tained using hybrid PET-MRI systems, which may be in-
markers. Combined with machine learning, such ap- formative when conventional radiology is negative [17,
proaches hold promise to match or exceed the accuracy 18]. Notwithstanding the practical advantages of this
of the evaluation by human experts [7]. technique, the added clinical value compared to tradition-
al single acquisitions with subsequent co-registration re-
mains to be established.
Hardware and Acquisition Techniques Currently, many epilepsy centers rely on 3T MRI for
routine clinical diagnostics. While the move to this plat-
Despite its unmatched diagnostic power, practices on form has significantly improved our ability to detect epi-
the use of MRI are still variable worldwide and do not har- leptogenic lesions [19], it is expected that resolving corti-
ness the full potential of technological advances for the cal laminar structures at 7T will likely push a step further
benefit of people with epilepsy. To standardize best-prac- our capabilities [20]. However, so far, it has been infre-
tice neuroimaging in outpatient clinics and specialized quent to see cortical dysplasias at 7T that are completely
surgery centers, the most recent guidelines propose the invisible at 3T [21, 22]. Also, for structural imaging of the
Harmonized Neuroimaging of Epilepsy Structural Se- neocortex and the immediate subcortical white matter,
quences (HARNESS) protocol, which includes easy to im- signal inhomogeneities particularly in the antero-inferior
plement, high-resolution 3D T1-weighted MRI, 3D temporal and frontal lobes still pose a challenge for the
FLAIR, and 2D coronal T2-weighted MRI [8]. Best qual- visualization of subtle lesions [20]. Conversely, the unique
ity for these widely available sequences is achieved at 3 possibility to perform molecular imaging of neurotrans-
Tesla; the use of multiple head coils allows for shorter mitters at 7T, such as GABA and glutamate [23], may
scanning time in addition to increased signal- and con- open new investigative avenues. Moreover, functional
trast-to-noise ratios. Notably, shorter acquisition times imaging is likely to reveal previously unresolved organi-
give the option to obtain additional contrasts to interro- zational features at a laminar level [24].

334 Eur Neurol 2022;85:333–341 Bernasconi/Bernasconi

DOI: 10.1159/000525262
 T1-weighted MRI T2-weighted MRI FLAIR/T1

Coronal
sections

+5
Sub
MRI-derived CA-3
features
CA4-DG
-5
z-score Volume Intensity FLAIR/T1

Fig. 1. Automated lateralization of “MRI-negative” HS. Lateraliza- for columnar volume, T2-weighted, and FLAIR/T1 intensities. On
tion prediction in a patient with MRI-negative left TLE using a each map (Sub = subiculum; CA1-3 = hippocampal CA1-3 sub-
linear discriminant classifier trained on T1- and FLAIR-derived fields; CA4-DG = hippocampal CA4 subfield and dentate gyrus),
laminar features of HS. As HS is typically characterized by T1- the dotted line corresponds to the level of the coronal MRI section,
weighted hypointensity and T2-weighted hyperintensity, the syn- and the optimal ROI obtained during training is outlined in black.

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

thetic contrast FLAIR/T1 maximizes their combined contribu- Histopathological analysis of the surgical specimen revealed subtle
tions to detect the full pathology spectrum. Coronal sections are HS mainly characterized by astrogliosis.
shown together with the automatically generated asymmetry maps

Lesion Detection rithm, operating on T1-weighted MRI, combines multi-

ple templates, parametric surfaces, and patch-based sam-
The essential role of structural lesions in the surgical pling for compact representation of image shape, texture,
management of drug-resistant epilepsy has motivated the and intensity, with unparalleled segmentation accuracy
development of increasingly sophisticated detection [39]. The combination of SurfPatch with 3D surface-
methods. This has been particularly relevant in patients based shape models [40, 41] sampling multicontrast fea-
with unremarkable routine MRI. Automated detection is tures along the central path of hippocampal subfields al-
generally performed by supervised classifiers trained to lows mapping morphological changes at a laminar level
learn imaging features that distinguish lesional from non- that may not be identified visually, thus furthering our
lesional. These image analysis techniques provide distinct knowledge of the HS spectrum [42]. Recently, a linear
information through quantitative assessment without the discriminant classifier trained on T1- and FLAIR-derived
cost of additional scanning time. laminar features of histologically validated HS accurately
In TLE, hippocampal sclerosis (HS) on MRI appears lateralized the focus in 93% of TLE patients, with a re-
as atrophy [25] and increased signal intensity [26], gener- markable 85% sensitivity in MRI-negative cases [43]
ally more severe ipsilateral to the seizure focus; accurate (Fig. 1). Notably, similar high performance in two inde-
identification of these signs is crucial for deciding the side pendent validation cohorts imaged on different scanners
of surgery. Whole-hippocampal volumetry has been one establishes generalizability, setting the basis for broad
of the first computational analyses applied to TLE [27– clinical translation.
31]. Advances in hardware and sequence technology, Among FCDs, encompassing various epileptogenic
which enable submillimetric resolution and an improved developmental malformations, type II lesions are the
signal-to-noise ratio, have facilitated accurate visualiza- most common and reliably diagnosed forms [44]. Histo-
tion of hippocampal subfields or subregions, including pathologically, FCD type II is typified by intracortical
the dentate gyrus, subiculum, and the cornu ammonis dyslamination and dysmorphic neurons, either in isola-
(CA1-4) regions [32]. Increasing demand to study large tion (IIA) or together with balloon cells (IIB). Conversely,
patient cohorts has motivated the shift from manual to- FCD type I remains elusive on histology and difficult to
ward automated segmentation, setting the basis for large- differentiate from normal cortex [45]. On MRI, while a
scale clinical use [33, 34]. Several methods with fast infer- previous study has shown cortical thinning at a group lev-
ence times have been developed for MRI-based subfield el, this malformation has been seen only in a handful of
segmentation [34–39], providing overlap indices of >80% pediatric cases [46]. FCD type II presents with an MRI
with manual labels. Among them, the SurfPatch algo- spectrum encompassing variable degrees of gray matter

MRI of Epilepsy Eur Neurol 2022;85:333–341 335

DOI: 10.1159/000525262
 probability | confidence
1.0
0.8 1
0.6 1
0.4 2 3 1
2
0.2
0
1 2 3 4 5 6 7 8 9 10
probability
Lesion FP cluster
low high

probability | confidence
1.0
0.8
0.6 3
3
0.4 5
1
0.2
0 1
1 2 3 4 5 6 7 8 9 10
rank (by degree of confidence)

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

Fig. 2. Automated “MRI-negative” FCD using deep learning. Two ter. The FCD lesion has both the highest probability (close to 1.0)
examples of FCD lesions initially reported as MRI-negative on and the highest confidence (rank 1). The right panels show the bi-
routine imaging. The left panels show the MRI sections and the nary maps of the FCD lesion (1) and FPs. Histopathological analy-
probability maps of putative FCD lesions as identified by the con- sis of the surgical specimens revealed a subtle FCD type IIA (dys-
volutional neural network. The bar plots show the probability of morphic neurons without balloon cells) in both cases. FP, false-
the lesion (purple) and FP (blue) clusters sorted by their rank; the positive.
superimposed line indicates the degree of confidence for each clus-

and white matter changes that challenges visual identifi- from relatively high false-positive rates [60]; also, the lim-
cation. Early approaches for computer-assisted visual de- ited set of features designed by human experts may not
tection were based on generic voxel-based morphometry, capture the full complexity of pathology. Alternatively, in
which demonstrated high sensitivity in detecting con- recent years, deep learning has shown high detection per-
spicuous malformations [47–50]; however, they failed to formance relative to conventional methods [64, 65]. In
identify two-thirds of subtle FCD lesions overlooked on particular, convolutional neural networks learn abstract
routine MRI [48, 50]. On the other hand, modeling of concepts from high-dimensional data, alleviating the
gray-white matter blurring and gray matter intensity challenging task of hand-crafting features [66]. In epilep-
[51], a simple method applicable to widely available 3D sy, a recent study applying deep learning to FCD detec-
T1-MRI clinical data, has been shown to be a very power- tion has reached a sensitivity of 93%, while maintaining
ful tool to assists visual inspection. Indeed, this approach high specificity both in healthy and disease controls [67].
increases sensitivity for the detection of FCD type II in at Importantly, this algorithm detected MRI-negative FCD
least 40% patients relative to the inspection of conven- type II with 85% sensitivity, thus offering a considerable
tional images. gain over standard radiological assessment. The algo-
Over the last decades, several surface-based algorithms rithm relied on T1- and T2-weighted FLAIR MRI of a
have been developed for fully automated detection of large cohort of patients with histologically validated le-
FCD type II [52–60]; the addition of FLAIR has increased sions collated across multiple tertiary centers. Results
sensitivity for the identification of smaller lesions [55]. were generalizable across cohorts with variable age, hard-
Importantly, careful preprocessing, including manual ware, and sequence parameters. A unique feature of this
corrections of tissue segmentation and surface extraction, algorithm was the use of Bayesian uncertainty estimation
have delivered high-fidelity FCD features [61]. Converse- for risk stratification [68, 69], which allowed to stratify
ly, suboptimal processing of surface-based data may lead predictions according to the confidence to be truly lesion-
to poor performance, even in cases with MRI visible le- al. In other words, putative lesional clusters were ranked
sions [62]. Admittedly, current benchmark automated based on confidence, thus assisting the examiner to gauge
detection algorithms fail in 20–40% of patients [55, 56, 59, the significance of findings (Fig. 2). By pairing predic-
63], particularly those with subtle FCD type II, and suffer tions with risk stratification, the classifier may assist clini-

336 Eur Neurol 2022;85:333–341 Bernasconi/Bernasconi

DOI: 10.1159/000525262
cians to adjust hypotheses relative to other tests, thereby Dirichlet allocation, an unsupervised topic modeling
increasing diagnostic confidence. technique, has significantly improved the prediction of
outcomes in the expression of disease factors. A recent
study in TLE applying this technique to multimodal MRI
Image-Based Prediction of Clinical Outcomes features of hippocampal and whole-brain gray and white
matter pathology uncovered specific dimensions of het-
Predicting seizure outcome after surgery has been ex- erogeneity not expressed in healthy controls and only
tensively explored in TLE. While the presence of ipsilat- minimally in patients with frontal lobe epilepsy [82]. Im-
eral hippocampal atrophy is a reliable indicator of seizure portantly, classifiers trained on the patients’ factor com-
freedom, alterations beyond the mesial temporal lobe position predicted response to antiseizure medications
may contribute to seizure relapse [70]. White matter mi- and surgery with an accuracy of 76% and 88%, respec-
crostructural features derived from DTI achieve high sen- tively, as well as memory scores, outperforming learners
sitivity but modest specificity [71, 72], similar to func- trained on group-level data (Fig. 3). In translational
tional connectivity measures of the thalamus and whole- terms, assessing interindividual variability identifies clin-
brain [73, 74]. While topological features of the ically relevant disease characteristics that would other-

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

structural connectome have shown predictive value for wise be missed. Integrating biotyping techniques that ex-
favorable outcome, their specificity for seizure relapse is ploit intra- and intersubject variability with other bio-
relatively low [75, 76]. Overall, paucity of large-scale val- markers, such as genomics, will likely offer novel avenues
idation and relatively low performance need to be ad- to elucidate disease processes at a molecular level [83].
dressed before advocating widespread clinical use of these
methods.
Neurological conditions, including epilepsy, are now- A Much-Needed Cultural Change Is Underway
adays conceptualized as heterogeneous disorders. This
has warranted the development of methods to explicitly Currently, many advances in neuroimaging of epilep-
model phenotypic variations across subjects, which may sy have not fully translated into clinical care. Indeed,
be ultimately exploited to predict outcomes [77]. Within practices on the use of imaging is variable worldwide as
biotyping techniques, categorical models provide sub- technical infrastructure and specialized training may not
types of patients with a given phenotype. In TLE, k-means be available or not sufficiently valued. The most dramat-
clustering of surface-based morphometry data uncovered ic implication of this translational gap relates to lesion
classes with distinct patterns of mesiotemporal anomalies identification, with many MRI-positive patients incor-
that differed with respect to histopathology and postsur- rectly labeled as MRI-negative. Paradoxically, while ap-
gical seizure outcome [42]. In FCD, leveraging hierarchi- pending MRI-negative status has profound implications
cal clustering to model connectivity from the lesion to the in terms of treatment strategies and seizure outcome, this
rest of the cortex can effectively predict surgical outcome crucial categorization lacks too often objectivity and rig-
[78]. In addition, a recent work based on consensus clus- or. This limitation has resulted in a plethora of SEEG, an
tering applied to multicontrast 3T MRI uncovered FCD invasive technique now accessible to many epilepsy cen-
tissue classes with distinct profiles of gray and white mat- ters, with the appealing yet debatable assumption that
ter anomalies, variably expressed within and across pa- electro-clinical hypotheses alone may be sufficient to
tients [79]; clinical utility was supported by gain in per- identify the epileptogenic zone. Importantly, SEEG is
formance of a lesion detection algorithm trained on class- nonlocalizing in more than 40% of patients, with the con-
informed data compared to class-naive paradigm. sequence that these cases will not undergo epilepsy sur-
Analyzing individual variability through unsupervised gery [84]. The remedy to this therapeutic dead-end is em-
techniques identified cognitive phenotypes associated bodied by the ongoing cultural shift fostered by several
with distinct patterns of white matter damage [80] and educational initiatives promoting neuroimaging skills,
connectome disorganization [81]. Compared to these the availability of standardized MRI acquisition protocols
categorical models, dimensional approaches allow a more and guidelines, as well as large-scale efforts within the
in-depth conceptualization of interindividual heteroge- neuroscience community to facilitate access to image
neity by uncovering axes of pathology that are co-ex- analyses techniques. Strengthening and widening the
pressed, to varying degrees, within and between individ- core competences of epileptologists will undoubtedly
ual patients. In neurodegenerative disorders, the latent transform traditional clinical decision-making into a

MRI of Epilepsy Eur Neurol 2022;85:333–341 337

DOI: 10.1159/000525262
 Factor-1 Factor-2
Neocortex Hippocampus

Atrophy

FLAIR hyperintensity

T1w/FLAIR decrease
White matter

FA decrease

MD increase
lpsi Contra lpsi Contra
Factor-3 Factor-4

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

Drug response Postsurgical seizure outcome

76±2.6% 88±1.5%
80 90

80
Balanced accuracy, %

70

70

60
60

50 50
Disease factors Disease factors
R

IR

FA

IR

FA

D
O

AI

AI
M

M
LA

LA
M

M
FL

FL
/F

/F
W

W
T1

T1

Fig. 3. Latent disease factors analysis in TLE. Maps show latent expression. High probability (darker red) signifies greater contri-
disease factors mapped onto neocortical, white matter, and hip- bution of a given feature to the factor (or high disease load). Drug
pocampal surfaces ipsi- and contralateral to the seizure focus. response and the seizure outcome after surgery are more accu-
These maps are derived from the application of latent Dirichlet rately predicted when using latent disease factors than when rely-
allocation, an unsupervised machine learning technique, to multi- ing on conventional group-level features. Data points indicate
modal MRI (T1w, FLAIR, T1w/FLAIR, diffusion-derived FA, and mean balanced accuracy for categorical data (drug response, sei-
MD) modeling atrophy, gliosis, demyelination, and microstruc- zure outcome) evaluated based on 100 repetitions of 10-fold cross-
tural damage. The latent Dirichlet allocation uncovers latent rela- validation.
tions (disease factors) from these features and quantifies their co-

338 Eur Neurol 2022;85:333–341 Bernasconi/Bernasconi

DOI: 10.1159/000525262
modern, systematic, multidisciplinary approach, ulti- Conflict of Interest Statement
mately circumventing the use of invasive diagnostic
The authors have no conflicts of interest to declare.
methods.

Funding Sources
Conclusion
This work was supported by the Canadian Institutes of Health
MRI is a mandatory investigation for diagnosis and Research to A.B. and N.B. (CIHR, MOP-57840 and 123520), the
treatment of epilepsy. The relentless progress in imaging Natural Sciences and Engineering Research Council of Canada
and machine learning techniques will continue to push (NSERC; Discovery-243141 to A.B. and 24779 to N.B.), and the
Epilepsy Canada Jay & Aiden Barker Breakthrough Grant in Clin-
the boundaries of lesion visibility and to provide increas- ical & Basic Sciences to A.B.
ingly sophisticated predictors of clinical outcomes for the
benefit of people with epilepsy.
This study is published in celebration of the 125th An- Author Contributions
niversary of the inception of European Neurology, 1897–

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

2022. Andrea Bernasconi and Neda Bernasconi: writing and editing
manuscript.

References
1 Wiebe S, Jette N. Pharmacoresistance and the 10 Tuch DS, Reese TG, Wiegell MR, Makris N, imaging-based multiparametric imaging in
role of surgery in difficult to treat epilepsy. Belliveau JW, Wedeen VJ. High angular reso- localizing seizure focus in patients with MRI-
Nat Rev Neurol. 2012;8:669. lution diffusion imaging reveals intravoxel negative temporal lobe epilepsy. AJNR Am J
2 Caciagli L, Bernasconi A, Wiebe S, Koepp MJ, white matter fiber heterogeneity. Magn Reson Neuroradiol. 2018;39(10):1791–8.
Bernasconi N, Bernhardt BC. A meta-analysis Med. 2002;48(4):577–82. 19 Knake S, Triantafyllou C, Wald LL, Wiggins
on progressive atrophy in intractable tempo- 11 Jones DK, Cercignani M. Twenty-five pitfalls G, Kirk GP, Larsson PG, et al. 3T phased array
ral lobe epilepsy: time is brain? Neurology. in the analysis of diffusion MRI data. NMR MRI improves the presurgical evaluation in
2017;89(5):506–16. Biomed. 2010;23(7):803–20. focal epilepsies: a prospective study. Neurol-
3 Keezer MR, Sisodiya SM, Sander JW. Comor- 12 Vargova L, Homola A, Cicanic M, Kuncova ogy. 2005;65(7):1026–31.
bidities of epilepsy: current concepts and fu- K, Krsek P, Marusic P, et al. The diffusion pa- 20 Verma G, Delman BN, Balchandani P. Ultra-
ture perspectives. Lancet Neurol. 2016;15(1): rameters of the extracellular space are altered High field MR imaging in epilepsy. Magn Re-
106–15. in focal cortical dysplasias. Neurosci Lett. son Imaging Clin N Am. 2021;29(1):41–52.
4 Jobst BC, Cascino GD. Resective epilepsy sur- 2011;499(1):19–23. 21 Veersema TJ, Ferrier CH, van Eijsden P,
gery for drug-resistant focal epilepsy: a re- 13 Tournier JD, Mori S, Leemans A. Diffusion Gosselaar PH, Aronica E, Visser F, et al. Sev-
view. JAMA. 2015;313(3):285–93. tensor imaging and beyond. Magn Reson en Tesla MRI improves detection of focal
5 Téllez-Zenteno JF, Ronquillo LH, Moien-Af- Med. 2011;65(6):1532–56. cortical dysplasia in patients with refractory
shari F, Wiebe S. Surgical outcomes in lesion- 14 Zhang H, Schneider T, Wheeler-Kingshott focal epilepsy. Epilepsia Open. 2017; 2(2):
al and non-lesional epilepsy: a systematic re- CA, Alexander DC. NODDI: practical in vivo 162–71.
view and meta-analysis. Epilepsy Res. 2010; neurite orientation dispersion and density 22 Feldman RE, Delman BN, Pawha PS, Dyvorne
89(2):310–8. imaging of the human brain. Neuroimage. H, Rutland JW, Yoo J, et al. 7T MRI in epi-
6 West S, Nolan SJ, Cotton J, Gandhi S, Weston 2012;61(4):1000–16. lepsy patients with previously normal clinical
J, Sudan A, et al. Surgery for epilepsy. Co- 15 Winston GP, Vos SB, Caldairou B, Hong SJ, MRI exams compared against healthy con-
chrane Database Syst Rev. 2015;7. Czech M, Wood TC, et al. Microstructural trols. PLoS One. 2019;14(3):e0213642.
7 Gabrieli John DE, Ghosh Satrajit S, Whitfield- imaging in temporal lobe epilepsy: diffusion 23 Davis KA, Nanga RP, Das S, Chen SH, Hadar
Gabrieli S. Prediction as a humanitarian and imaging changes relate to reduced neurite PN, Pollard JR, et al. Glutamate imaging (Glu-
pragmatic contribution from human cogni- density. Neuroimage Clin. 2020;26:102231. CEST) lateralizes epileptic foci in nonlesional
tive neuroscience. Neuron. 2015;85(1):11–26. 16 Bernhardt BC, Fadaie F, Vos de Wael R, Hong temporal lobe epilepsy. Sci Transl Med. 2015;
8 Bernasconi A, Cendes F, Theodore WH, Gill SJ, Liu M, Guiot MC, et al. Preferential sus- 7(309):309ra161.
RS, Koepp MJ, Hogan RE, et al. Recommen- ceptibility of limbic cortices to microstructur- 24 Morris LS, Kundu P, Costi S, Collins A,
dations for the use of structural magnetic res- al damage in temporal lobe epilepsy: a quan- Schneider M, Verma G, et al. Ultra-high field
onance imaging in the care of patients with titative T1 mapping study. Neuroimage. 2018 MRI reveals mood-related circuit disturbanc-
epilepsy: a consensus report from the Interna- Nov 15;182:294–303. es in depression: a comparison between
tional League Against Epilepsy Neuroimag- 17 Traub-Weidinger T, Muzik O, Sundar LKS, 3-Tesla and 7-Tesla. Transl Psychiatry. 2019;
ing Task Force. Epilepsia. 2019; 60(6): 1054– Aull-Watschinger S, Beyer T, Hacker M, et al. 9(1):94.
68. Utility of absolute quantification in non-le- 25 Berkovic SF, Andermann F, Olivier A, Ethier
9 Deleo F, Thom M, Concha L, Bernasconi A, sional extratemporal lobe epilepsy using FDG R, Melanson D, Robitaille Y, et al. Hippocam-
Bernhardt BC, Bernasconi N. Histological PET/MR imaging. Front Neurol. 2020;11:54. pal sclerosis in temporal lobe epilepsy dem-
and MRI markers of white matter damage in 18 Shang K, Wang J, Fan X, Cui B, Ma J, Yang H, onstrated by magnetic resonance imaging.
focal epilepsy. Epilepsy Res. 2017;140:29–38. et al. Clinical value of hybrid TOF-PET/MR Ann Neurol. 1991;29(2):175–82.

MRI of Epilepsy Eur Neurol 2022;85:333–341 339

DOI: 10.1159/000525262
26 Kuzniecky R, de la Sayette V, Ethier R, Melan- 38 Kim H, Bernhardt BC, Kulaga-Yoskovitz J, 50 Salmenpera TM, Symms MR, Rugg-Gunn FJ,
son D, Andermann F, Berkovic S, et al. Mag- Caldairou B, Bernasconi A, Bernasconi N. Boulby PA, Free SL, Barker GJ, et al. Evalua-
netic resonance imaging in temporal lobe ep- Multivariate hippocampal subfield analysis of tion of quantitative magnetic resonance im-
ilepsy: pathological correlations. Ann Neurol. local MRI intensity and volume: application aging contrasts in MRI-negative refractory
1987;22(3):341–7. to temporal lobe epilepsy. Med Image Com- focal epilepsy. Epilepsia. 2007;48(2):229–37.
27 Cascino GD, Jack CR Jr, Parisi JE, Sharbrough put Comput Assist Interv. 2014;17:170–8. 51 Bernasconi A, Antel SB, Collins DL, Ber-
FW, Hirschorn KA, Meyer FB, et al. Magnet- 39 Caldairou B, Bernhardt BC, Kulaga-Yoskov- nasconi N, Olivier A, Dubeau F, et al. Texture
ic resonance imaging-based volume studies in itz J, Kim H, Bernasconi N, Bernasconi A. A analysis and morphological processing of
temporal lobe epilepsy: pathological correla- surface patch-based segmentation method for magnetic resonance imaging assist detection
tions. Ann Neurol. 1991;30(1):31–6. hippocampal subfields. MICCAI 2016, the of focal cortical dysplasia in extra-temporal
28 Cendes F, Andermann F, Gloor P, Evans A, 19th International Conference on Medical partial epilepsy. Ann Neurol. 2001;49(6):770–
Jones-Gotman M, Watson C, et al. MRI volu- Image Computing and Computer Assisted 5.
metric measurement of amygdala and hippo- Intervention. 2016. Vol. 9901; p. 379–87. Lec- 52 Adler S, Wagstyl K, Gunny R, Ronan L, Car-
campus in temporal lobe epilepsy. Neurology. ture notes in computer science. michael D, Cross JH, et al. Novel surface fea-
1993;43(4):719. 40 Hogan RE, Wang L, Bertrand ME, Willmore tures for automated detection of focal cortical
29 Watson C, Andermann F, Gloor P, Jones- LJ, Bucholz RD, Nassif AS, et al. Predictive dysplasias in paediatric epilepsy. Neuroimage
Gotman M, Peters T, Evans A, et al. Anatom- value of hippocampal MR imaging-based Clin. 2016;14:18–27.
ic basis of amygdaloid and hippocampal vol- high-dimensional mapping in mesial tempo- 53 Ahmed B, Brodley CE, Blackmon KE,
ume measurement by magnetic resonance ral epilepsy: preliminary findings. AJNR Am Kuzniecky R, Barash G, Carlson C, et al. Cor-
imaging. Neurology. 1992;42(9):1743. J Neuroradiol. 2006;27(10):2149–54. tical feature analysis and machine learning

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

30 Watson C, Jack CR, Cendes F. Volumetric 41 Posener JA, Wang L, Price JL, Gado MH, improves detection of “MRI-negative” focal
magnetic resonance imaging: clinical applica- Province MA, Miller MI, et al. High-dimen- cortical dysplasia. Epilepsy Behav. 2015; 48:
tions and contributions to the understanding sional mapping of the hippocampus in de- 21–8.
of temporal lobe epilepsy. Arch Neurol. 1997; pression. Am J Psychiatry. 2003;160(1):83–9. 54 Ahmed B, Thesen T, Blackmon KE, Kuzniek-
54(12):1521–31. 42 Bernhardt BC, Hong SJ, Bernasconi A, Ber- cy R, Devinsky O, Brodley CE. Decrypting
31 Bernasconi N, Bernasconi A, Caramanos Z, nasconi N. Magnetic resonance imaging pat- “cryptogenic” epilepsy: semi-supervised hier-
Antel SB, Andermann F, Arnold DL. Mesial tern learning in temporal lobe epilepsy: clas- archical conditional random fields for detect-
temporal damage in temporal lobe epilepsy: a sification and prognostics. Ann Neurol. 2015; ing cortical lesions in MRI-negative patients.
volumetric MRI study of the hippocampus, 77(3):436–46. J Mach Learn Res. 2016;17(1):3885–914.
amygdala and parahippocampal region. 43 Caldairou B, Foit NA, Mutti C, Fadaie F, Gill 55 Gill RS, Hong S-J, Fadaie F, Caldairou B, Ber-
Brain. 2003;126(2):462–9. R, Lee HM, et al. An MRI-based machine nhardt B, Bernasconi N, et al. Automated de-
32 Kulaga-Yoskovitz J, Bernhardt BC, Hong S-J, learning prediction framework to lateralize tection of epileptogenic cortical malforma-
Mansi T, Liang KE, van der Kouwe AJW, et al. hippocampal sclerosis in patients with tem- tions using multimodal MRI. Deep Learning
Multi-contrast submillimetric 3 Tesla hippo- poral lobe epilepsy. Neurology. 2021; 97(16): in Medical Image Analysis and Multimodal
campal subfield segmentation protocol and e1583–93. Learning for Clinical Decision Support: Third
dataset. Sci Data. 2015;2:150059. 44 Blumcke I, Thom M, Aronica E, Armstrong International Workshop, DLMIA 2017, ML-
33 Shishegar R, Gandomkar Z, Soltaman-Zadeh DD, Vinters HV, Palmini A, et al. The clini- CDS 2017. 2017. Vol. 10553; p. 349–56. Lec-
H, Moghadasi SR. Hippocampal shape analy- copathologic spectrum of focal cortical dys- ture notes in computer science.
sis in the Laplace Beltrami feature space for plasias: a consensus classification proposed 56 Hong SJ, Kim H, Schrader D, Bernasconi N,
temporal lobe epilepsy diagnosis and lateral- by an ad hoc Task Force of the ILAE Diagnos- Bernhardt BC, Bernasconi A. Automated de-
ization. 2012 9th IEEE International Sympo- tic Methods Commission. Epilepsia. 2011; tection of cortical dysplasia type II in MRI-
sium on Biomedical Imaging [ISBI]. 2012. p. 52(1):158–74. negative epilepsy. Neurology. 2014;83(1):48–
150–3. 45 Blumcke I, Coras R, Busch RM, Morita-Sher- 55.
34 Kim H, Mansi T, Bernasconi N, Bernasconi A. man M, Lal D, Prayson R, et al. Toward a bet- 57 Jin B, Krishnan B, Adler S, Wagstyl K, Hu W,
Robust surface-based multi-template auto- ter definition of focal cortical dysplasia: an it- Jones S, et al. Automated detection of focal
mated algorithm to segment healthy and erative histopathological and genetic agree- cortical dysplasia type II with surface-based
pathological hippocampi. Med Image Com- ment trial. Epilepsia. 2021;62(6):1416–28. magnetic resonance imaging postprocessing
put Comput Assist Interv. 2011;14:445–53. 46 Blumcke I, Pieper T, Pauli E, Hildebrandt M, and machine learning. Epilepsia. 2018; 59(5):
35 Iglesias JE, Augustinack JC, Nguyen K, Player Kudernatsch M, Winkler P, et al. A distinct 982–92.
CM, Player A, Wright M, et al. A computa- variant of focal cortical dysplasia type I char- 58 Tan Y-L, Kim H, Lee S, Tihan T, Ver Hoef L,
tional atlas of the hippocampal formation us- acterised by magnetic resonance imaging and Mueller SG, et al. Quantitative surface analy-
ing ex vivo, ultra-high resolution MRI: appli- neuropathological examination in children sis of combined MRI and PET enhances de-
cation to adaptive segmentation of in vivo with severe epilepsies. Epileptic Disord. 2010; tection of focal cortical dysplasias. Neuroim-
MRI. Neuroimage. 2015;115:117–37. 12(3):172–80. age. 2018;166:10–8.
36 Pipitone J, Park MTM, Winterburn J, Lett TA, 47 Focke NK, Symms MR, Burdett JL, Duncan 59 Snyder K, Whitehead EP, Theodore WH, Za-
Lerch JP, Pruessner JC, et al. Multi-atlas seg- JS. Voxel-based analysis of whole brain ghloul KA, Inati SJ, Inati SK. Distinguishing
mentation of the whole hippocampus and FLAIR at 3T detects focal cortical dysplasia. type II focal cortical dysplasias from normal
subfields using multiple automatically gener- Epilepsia. 2008;49(5):786–93. cortex: a novel normative modeling ap-
ated templates. Neuroimage. 2014; 101: 494– 48 Rugg-Gunn FJ, Eriksson SH, Boulby PA, proach. Neuroimage Clin. 2021;30:102565.
512. Symms MR, Barker GJ, Duncan JS. Magneti- 60 Kini LG, Gee JC, Litt B. Computational analy-
37 Van Leemput K, Bakkour A, Benner T, Wig- zation transfer imaging in focal epilepsy. Neu- sis in epilepsy neuroimaging: a survey of fea-
gins G, Wald LL, Augustinack J, et al. Auto- rology. 2003;60(10):1638–45. tures and methods. Neuroimage Clin. 2016;
mated segmentation of hippocampal sub- 49 Rugg-Gunn FJ, Boulby PA, Symms MR, Bark- 11:515–29.
fields from ultra-high resolution in vivo MRI. er GJ, Duncan JS. Whole-brain T2 mapping
Hippocampus. 2009;19(6):549–57. demonstrates occult abnormalities in focal
epilepsy. Neurology. 2005;64(2):318–25.

340 Eur Neurol 2022;85:333–341 Bernasconi/Bernasconi

DOI: 10.1159/000525262
61 Gill RS, Caldairou B, Bernasconi N, Bernasco- 69 Gal Y, Ghahramani Z. Bayesian convolution- 77 Arbabshirani MR, Plis S, Sui J, Calhoun VD.
ni A. Uncertainty-informed detection of epi- al neural networks with Bernoulli approxi- Single subject prediction of brain disorders in
leptogenic brain malformations using Bayes- mate variational inference. 2015. Epub ahead neuroimaging: promises and pitfalls. Neuro-
ian neural networks. International Confer- of print. image. 2017;145:137–65.
ence on Medical Image Computing and 70 Bernhardt BC, Hong S, Bernasconi A, Ber- 78 Hong S-J, Lee H-M, Gill R, Crane J, Sziklas V,
Computer-Assisted Intervention: MICCAI nasconi N. Imaging structural and functional Bernhardt BC, et al. A connectome-based
2019. 2019. Vol. 11767; p. 225–33. Lecture brain networks in temporal lobe epilepsy. mechanistic model of focal cortical dysplasia.
notes in computer science. Front Hum Neurosci. 2013;7(624):624. Brain. 2019;142(3):688–99.
62 Spitzer H, Ripart M, Whitaker K, Napolitano 71 Munsell BC, Wee C-Y, Keller SS, Weber B, 79 Lee HM, Gill RS, Fadaie F, Cho KH, Guiot
A, De Palma L, De Benedictis A, et al. Inter- Elger C, da Silva LAT, et al. Evaluation of ma- MC, Hong S-J, et al. Unsupervised machine
pretable surface-based detection of focal cor- chine learning algorithms for treatment out- learning reveals lesional variability in focal
tical dysplasias: a MELD study. medRxiv. come prediction in patients with epilepsy cortical dysplasia at mesoscopic scale. Neuro-
2021. based on structural connectome data. Neuro- image Clin. 2020;28:102438.
63 Zhao Y, Ahmed B, Thesen T, Blackmon KE, image. 2015;118:219–30. 80 Reyes A, Kaestner E, Bahrami N, Balachandra
Dy JG, Brodley CE, et al. A non-parametric 72 Taylor PN, Sinha N, Wang Y, Vos SB, De Tisi A, Hegde M, Paul BM, et al. Cognitive pheno-
approach to detect epileptogenic lesions using J, Miserocchi A, et al. The impact of epilepsy types in temporal lobe epilepsy are associated
restricted Boltzmann machines. Proceedings surgery on the structural connectome and its with distinct patterns of white matter network
of the 22nd ACM SIGKDD International relation to outcome. Neuroimage Clin. 2018; abnormalities. Neurology. 2019; 92(17):
Conference on Knowledge Discovery and 18:202–14. e1957–68.
Data Mining. 2016. p. 373–82. 73 He X, Doucet GE, Pustina D, Sperling MR, 81 Rodríguez-Cruces R, Bernhardt BC, Concha

Downloaded from http://karger.com/ene/article-pdf/85/5/333/3749765/000525262.pdf by guest on 09 January 2024

64 Litjens G, Kooi T, Bejnordi BE, Setio AAA, Sharan AD, Tracy JI. Presurgical thalamic L. Multidimensional associations between
Ciompi F, Ghafoorian M, et al. A survey on “hubness” predicts surgical outcome in tem- cognition and connectome organization in
deep learning in medical image analysis. Med poral lobe epilepsy. Neurology. 2017; 88(24): temporal lobe epilepsy. Neuroimage. 2020;
Image Anal. 2017;42:60–88. 2285–93. 213:116706.
65 Topol EJ. High-performance medicine: the 74 Larivière S, Weng Y, Vos de Wael R, Royer J, 82 Lee HM, Fadaie F, Gill R, Caldairou B, Sziklas
convergence of human and artificial intelli- Frauscher B, Wang Z, et al. Functional con- V, Crane J, et al. Decomposing MRI pheno-
gence. Nat Med. 2019;25(1):44–56. nectome contractions in temporal lobe epi- typic heterogeneity in epilepsy: a step towards
66 Goodfellow I, Bengio Y, Courville A. Deep lepsy: microstructural underpinnings and personalized classification. Brain. 2022 Apr
learning. MIT press; 2016. predictors of surgical outcome. Epilepsia. 29;145(3):897–908.
67 Gill RS, Lee HM, Caldairou B, Hong SJ, Barba 2020;61(6):1221–33. 83 Arnatkevic̆iūtė A, Fulcher BD, Fornito A. A
C, Deleo F, et al. Multicenter validation of a 75 Gleichgerrcht E, Keller SS, Drane DL, Mun- practical guide to linking brain-wide gene ex-
deep learning detection algorithm for Ffocal sell BC, Davis KA, Kaestner E, et al. Temporal pression and neuroimaging data. Neuroim-
cortical dysplasia. Neurology. 2021; 97(16): lobe epilepsy surgical outcomes can be in- age. 2019;189:353–67.
e1571–82. ferred based on structural connectome hubs: 84 Cardinale F, Rizzi M, Vignati E, Cossu M,
68 Leibig C, Allken V, Ayhan MS, Berens P, a machine learning study. Ann Neurol. 2020; Castana L, d’Orio P, et al. Stereoelectroen-
Wahl S. Leveraging uncertainty information 88(5):970–83. cephalography: retrospective analysis of 742
from deep neural networks for disease detec- 76 Sinha N, Wang Y, da Silva NM, Miserocchi A, procedures in a single centre. Brain. 2019;
tion. Sci Rep. 2017;7(1):17816. McEvoy AW, de Tisi J, et al. Structural brain 142(9):2688–704.
network abnormalities and the probability of
seizure recurrence after epilepsy surgery.
Neurology. 2021;96(5):e758–71.

MRI of Epilepsy Eur Neurol 2022;85:333–341 341

DOI: 10.1159/000525262