The Capabilities of Multimodal Magnetic Resonance Morphometry to Diagnose Complex Cases of Focal Cortical Dysplasia (Pilot Study) Authors: AM Shevchenko, NY Alsahanova, EL Pogosbekyan, AI Batalov, NE Elagina, AA Kopaneva, AV Bernstein, MG Sharaev, IN Pronin Published: 2026 ABSTRACT AbstractFocal cortical dysplasia (FCD) is a leading cause of pharmacoresistant epilepsy. Up to 30–50% of cases remain MRI-negative on routine visual assessment. MR morphometry improves dysplasia detection; however, conventional approaches are limited to the analysis of T1-weighted images and only two features (gray–white matter blurring and cortical thickness abnormalities).The aim of the study was to compare the diagnostic efficiency of the conventional T1-based two-feature algorithm and a multimodal multi-feature algorithm incorporating machine learning.Materials and MethodsThe study included four patients with histologically verified FCD (2 patients with type I and 2 patients with type IIB according to the Blümcke classification). All patients underwent MRI using the HARNESS protocol. MR morphometry was performed using two independent algorithms.The conventional method included FreeSurfer-based assessment of 2 features (the analysis of cortical thickening and gray–white matter boundary blurring) on T1-weighted images.The multimodal morphometric algorithm incorporated T1/T2-weighted images and FLAIR, and included the following features: gray–white matter transition abnormalities on T1, T2, and T2-FLAIR images (Blurring_T1/T2/FLAIR); detection of local T2- and FLAIR-hyperintense signal abnormalities (CR_T2/FLAIR); assessment of FLAIR intensity distribution heterogeneity to identify blurred boundaries or the transmantle sign (Variance); measurement of signal irregularity on FLAIR images to detect the regions of increased contrast (Entropy). Machine learning models integrating all extracted features were used to generate final FCD prediction maps (MLP and XGBoost).ResultsIn two cases, where FCD had typical T1-signs (cortical thickening or well-defined demarcation impairment), both algorithms demonstrated equal diagnostic performance. In the third case, the classical approach failed to detect pathology, since only T1-weighted images were analyzed, whereas the multimodal approach enabled to identify dysplastic areas (only T2/FLAIR images showed the changes). In the fourth case, the conventional algorithm detected a single lesion, while the multimodal algorithm identified an additional second lesion, which influenced the surgical treatment strategy.ConclusionMultimodal MR morphometry with artificial intelligence outperforms the conventional approach in complex cases (MRI-negative and multifocal lesions) and should be considered an important supplementary technique in preoperative planning.Keywords: focal cortical dysplasia, MR morphometry, artificial intelligence, machine learning, pharmacoresistant epilepsy, multimodal MRI Introduction Focal cortical dysplasia (FCD) is one of the most common causes of drug-resistant epilepsy, especially in children, and is a leading cause of structural epilepsy in candidates for surgical treatment [1–3]. The only treatment for such patients is radical surgical resection of the epileptogenic focus; however, the success of the surgery directly depends on the accurate preoperative determination of the localization of the boundaries of the dysplastic cortical area [1]. The gold standard for FCD diagnosis is visual assessment using high-resolution magnetic resonance imaging (MRI) according to specialized protocols, such as HARNESS. However, this approach has significant limitations. Dysplastic lesions are often very small, located deep within complex sulci and gyri, or show only subtle structural changes [3]. As a result, a substantial proportion of FCD cases (reported to be as high as 30–50%) remain undetected by routine visual analysis, contributing to what is known as MRI-negative epilepsy [1, 4]. Additionally, interpreting MRI results is subject to considerable inter-expert variability and requires significant time investment [3]. Because of these challenges, there is a strong need for objective and sensitive analytical methods capable of detecting hidden structural abnormalities. MRI morphometry based on voxel image analysis and data post-processing, offers a way to quantitatively assess key features of FCD, such as local cortical thickening and blurring of the gray–white matter boundary [1, 4]. Our studies have shown the use of MR morphometry to enable to increase the detection rate of FCD type I focal cortical dysplasia by 18% (up to 85%) compared to standard visual assessment (66%), which is of particular importance due to the challenges in diagnosing this category of dysplasia [1]. MR morphometry has been clinically proven to be valuable in identifying lesions in patients with prolonged MRI-negative epilepsy. Subsequent histological studies confirmed the presence of dysplasia. Postoperatively, these patients experienced sustained remission of epileptic seizures [4]. Currently, efforts are focused on improving the method specificity and reducing the number of false-positive results. Promising directions include integrating MR morphometry with quantitative MRI techniques, such as MR fingerprinting, as well as the development of machine learning algorithms to automatically distinguish true pathological lesions from normal variants [2, 3]. Moreover, the studies are underway to assess how the quality of original MR images impacts post-processing outcomes — the aspect is critical for the proper standardization of the method [2]. In summary, exploring the diagnostic potential of MR morphometry, optimizing its algorithms, and integrating it into clinical practice are highly relevant goals to enhance patient selection for surgical treatment and improving its overall effectiveness. The aim of the study was to compare the diagnostic efficiency of two MR morphology algorithms in patients with pharmacoresistant epilepsy suspicious for focal cortical dysplasia. The first algorithm (N.N. Burdenko National Medical Research Center of Neurosurgery) is a classical method based solely on the analysis of T1-weighted images, evaluating two features (cortical thickening and gray–white matter blurring). The second algorithm (BIMAI-Lab, Skolkovo Institute of Science and Technologies) is a multimodal method (T1, T2, FLAIR) with the calculation of an expanded set of morphometric features. Materials and Methods The prospective study included four patients diagnosed with FCD types I and II, all suffering from pharmacoresistant structural epilepsy (one patient with obvious FCD; one MRI-negative patient; one with FCD visible only on T2 and T2-FLAIR imaging; and one with extensive FCD). Prior to surgery, all patients underwent brain MRI following the HARNESS protocol using either a General Electric Signa HDxt (GE Healthcare, USA) 3.0 Tesla scanner with an 8-channel head coil or a MAGNETOM Skyra (Siemens, Germany) 3.0 Tesla scanner with a 32-channel head coil. The MRI images were analyzed by a multidisciplinary team consisting of radiologists, epileptologists, and neurosurgeons. The imaging findings, clinical data, and other diagnostic test results were presented at a multidisciplinary conference to determine the surgical strategy and scope aimed at resecting the epileptogenic zone. After surgery, FCD was confirmed in all patients; two were diagnosed with FCD type I, and two with FCD type IIB according to Blümcke classification. The study was conducted in compliance with the 2024 Declaration of Helsinki and approved by the Ethics Committee of the N.N. Burdenko National Medical Research Center of Neurosurgery (Russia). Multimodal MRI morphometry algorithm using machine learning (BIMAI-Lab, Skolkovo Institute of Science and Technologies) Within the study framework, a set of quantitative feature maps associated with FCD were generated and normalized against a “healthy” database to improve the sensitivity to subtle changes including MRI-negative cases. The following morphometric features were computed: Blurring_T1: a feature aimed at detecting blurred boundaries between gray and white matter measured on T1-weighted MRI; Blurring_T2: a feature detecting blurred gray–white matter boundaries measured on T2-weighted MRI; Blurring_FLAIR: a feature detecting blurred gray–white matter boundaries measured on FLAIR MRI; CR_T2: a feature aimed at detecting an increased signal (the gray–white matter hyperintensity boundary) measured on T2-weighted MRI. CR (concentration rate) indicates abnormally high signal intensity in the area, and this feature typically detects hyperintensities well; CR_FLAIR is a feature aimed at detecting an increased signal (the hyperintensity boundary between gray/white matter) measured by FLAIR MRI. CR indicates that the signal is abnormally high for the area. This feature typically detects hyperintensity well; Variance is a feature that helps detect blurriness at the interface between gray and white matter or a transmantle sign measured by FLAIR MRI. This feature is more related to the irregular distribution of intensity in the area; i.e. blurring at the boundary or a transmantle sign in white matter leads to an increase in the value of Variance; Entropy is a feature that enables to identify the areas of increased signal (hyperintensity boundaries of gray matter/white matter) and a transmantle sign characterized by higher contrast between the area of interest and its surroundings. This feature is calculated using FLAIR and reflects the randomness of the signal; MLP, XGBoost are the models built on features (they generalize the information from the features and provide predictions). The output involved the formation of maps detailing morphometric and contrast features as well as prediction maps from the trained machine learning models (MLP and XGBoost [5]) [6]. The main stages of this algorithm are described below. Data preparation and spatial normalization The original 3D MRI images (T1, T2, FLAIR) were anonymized and converted to the NIfTI format. The T1 images were registered to the symmetrical standard MNI152 template space (based on 152 MRIs) [7]. T2 and FLAIR modalities were then transformed into MNI152 space using linear registration to ensure voxel-wise correspondence across the modalities. After registration, intensity normalization (z-score) and correction for MRI signal inhomogeneity were performed. Tissue segmentation and the creation of probabilistic maps of gray matter (GM), white matter (WM), and cerebrospinal fluid were conducted using SPM12 [8]; voxel values on each map represented the probability of belonging to the corresponding tissue type. Brain masks were generated using HD-BET [9]. Extraction of morphometric features from T1 Morphometric maps of cortical thickness, curvature, and sulcal depth were calculated from T1 images using FreeSurfer (recon-all) [10]. To refine cortical boundaries, the FLAIR modality was used as an auxiliary input. The resulting surface maps were converted into volumetric maps and registered to the MNI152 template space for subsequent voxel-wise analysis. Boundary blurring maps between gray and white matter for T1, T2, and FLAIR For each modality, a “gray–white matter” boundary map was created based on individual thresholds derived from intensity distributions within gray matter (GM) and white matter (WM). For T1, the thresholds were defined as: where μ and σ represent the mean and standard deviation of intensities within the corresponding tissue types, and δ is an empirical parameter adjusted to satisfy the condition Tlow