Authors :
Presenting Author: Ting-Yu Su, PhD – Cleveland Clinic
Siyuan Hu, PhD – Case Western Reserve University
Xiaofeng Wang, PhD – Cleveland Clinic
Sophie Adler, PhD – University College London Great Ormond Street Institute of Child Health
Konrad Wagstyl, PhD – University College London Great Ormond Street Institute of Child Health
Zheng Ding, PhD – Cleveland Clinic
Spencer Morris, MS – Cleveland Clinic
Joon Yul Choi, PhD – Cleveland Clinic
Ken Sakaie, PhD – Cleveland Clinic
Ingmar Blümcke, MD – University Hospitals Erlangen
Hiroatsu Murakami, MD, PhD – Cleveland Clinic
Stephen Jones, MD – Cleveland Clinic
Imad Najm, MD – Cleveland Clinic
Dan ma, PhD – Duke University
Zhong Irene Wang, PhD – Cleveland Clinic
Rationale:
Focal cortical dysplasia (FCD) is a prevalent cause of drug-resistant focal epilepsy, yet subtle lesions often remain undetected on standard MRI. Magnetic resonance fingerprinting (MRF) is an innovative imaging modality capable of simultaneously producing quantitative tissue maps within a single rapid scan. This study proposes a fully automated detection pipeline leveraging surface-based morphometric (SBM) analysis of high-resolution MRF data, aiming to improve FCD identification and to reduce false positives (FPs).Methods:
The study involved 114 participants, including 44 patients with drug-resistant focal epilepsy and FCD as well as 70 healthy controls (HCs). All subjects underwent whole-brain 3T MRF scanning to obtain quantitative T1 and T2 maps, along with clinical T1-weighted (T1w) images. 35 patients also had 3D fluid attenuated inversion recovery (FLAIR) data. Expert-defined lesion masks were manually delineated and registered to the T1w image. Using the Multi-center Epilepsy Lesion Detection pipeline, surface-based features were extracted from the T1w images. The MRF T1 and T2 maps, along with FLAIR images, were registered to the Freesurfer domain and sampled at multiple cortical depths (at 25%, 50%, and 75% of the cortical thickness, the gray-white matter interface, as well as 0.5 mm and 1 mm into the white matter). Feature normalization was performed using intra-subject, inter-hemispheric, and inter-subject z-scoring. We used a two-stage machine learning (ML) pipeline: (1) a vertex-level neural network classifier trained to distinguish lesional from normal vertices, and (2) a secondary RUSBoost classifier operating at the cluster level to reduce FPs based on cluster size, vertex-wise probability, and feature summary metrics. Leave-one-out cross-validation was performed at both stages. Figure 1 details the study workflow.Results:
Using T1w images alone yielded a sensitivity of 70.4% for individual-level lesion detection, but with 11.6 FP clusters per patient and 4.1 per HC. Incorporation of MRF features lowered FP clusters to 6.6 and 1.5 per patient and control, respectively, with a modest drop in sensitivity to 68.2%. When T1w, MRF, and FLAIR were combined, sensitivity improved to 71.4% while reducing FPs further (4.7 in patients, 1.1 in controls). Detection rates were higher in type II FCD cases compared to non-type II, and MRI-positive cases outperformed MRI-negative ones. Furthermore, patients who were seizure-free postoperatively had greater lesion detection rates. Subtype classification yielded 80.8% accuracy between non-type II and type II cases, and 68.4% between IIa and IIb. Notably, the transmantle sign was observed in 62.5% of IIb and 40% of IIa cases. Figure 2 shows lesion detection for 8 example patients.Conclusions:
We established a surface-based MRF pipeline with promising efficacy for FCD detection/subtyping, achieving FP suppression through a two-stage classification. Selected cluster metrics may serve as indicators of detection confidence and seizure outcomes.
Funding:
NIH R01 NS109439, 2R01 NS109439