Automated Surfaced-based Detection of Focal Cortical Dysplasia Using MR Fingerprinting
Abstract number :
3.246
Submission category :
5. Neuro Imaging / 5A. Structural Imaging
Year :
2022
Submission ID :
2204934
Source :
www.aesnet.org
Presentation date :
12/5/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:27 AM
Authors :
Ting-Yu Su, PhD – Cleveland Clinic; Siyuan Hu, PhD – Biomedical Engineering – Case Western Reserve University; Xiaofeng Wang, PhD – Quantitative Health Science – Cleveland Clinic; Sophie Adler, MB PhD – Great Ormond Street Hospital – National Health Service Foundation Trust; Konrad Wagstyl, MB PhD – Wellcome Centre for Human Neuroimaging; Zheng Ding, PhD – Charles Shor Epilepsy Center – Cleveland Clinic; Joon Yul Choi, PhD – Charles Shor Epilepsy Center – Cleveland Clinic; Ken Sakaie, PhD – Imaging Institute – Cleveland Clinic; Ingmar Blümcke, MD – Neuropathology – University Hospitals Erlangen; Hiroatsu Murakami, MD – Charles Shor Epilepsy Center – Cleveland Clinic; Stephen Jones, MD PhD – Imaging Institute – Cleveland Clinic; Imad Najm, MD – Charles Shor Epilepsy Center – Cleveland Clinic; Dan Ma, PhD – Biomedical Engineering – Case Western Reserve University; Zhong Irene Wang, PhD – Charles Shor Epilepsy Center – Cleveland Clinic
Rationale: Focal cortical dysplasia (FCD) is a common pathology in pharmacoresistant focal epilepsy. Subtle FCDs can be difficult to detect by visual inspection of conventional MRI. Magnetic resonance fingerprinting (MRF) is an advanced quantitative MR technique that can acquire multiple tissue property maps simultaneously within a clinically acceptable timespan. We aimed to develop a framework for automated FCD detection using surface-based morphometric processing of conventional MRI and MRF.
Methods: Thirty-six patients with FCD (histologically confirmed in 29, radiologically confirmed in 7) were included, as well as 48 healthy controls (HCs). MRI scans were performed using a Siemens 3T Prisma scanner. A 3D whole-brain MRF sequence (1 mm3 isotropic voxels) was used (Ma et al., JMRI 2019). Dictionary-based reconstruction of the MRF T1 and T2 maps was performed. A 3D ROI was created for each FCD lesion. We utilized the MELD pipeline (Wagstyl et al., Epilepsia 2021) to generate surface-based morphometric features on clinical T1w images, including thickness, mean curvature, intrinsic curvature, grey-white contrast, sulcal depth, and asymmetry maps of these features. The MRF T1 and T2 maps as well as 3D FLAIR (available in 29 of the 36 patients) were registered to the T1w images and sampled at 25%, 50%, and 75% of the cortical thickness, as well as at the gray‐white matter boundary and 0.5 and 1 mm subcortically. Normalization procedures with intra-subject, intra-hemispheric and inter-subject z-scoring using the HCs were performed. Uniform Manifold Approximation and Projection (UMAP) was used for visualization of features separating patients and HCs. Using different feature sets as input, neural network models with 2 hidden layers with 40 and 10 nodes were trained using patient and HC data. The prediction outcomes were binarized based on the optimal threshold determined by the Youden’s J index on the training cohort. Nested leave-one-out cross-validation was used to assess lesion detection performance. Receiver operating characteristic (ROC) plot was used to evaluate vertex-wise performance.
Results: UMAP results (Figure 1) showed MRF and FLAIR features both improved separation between patients and HCs, as compared to using T1w features alone. Vertex-wise performance (Figure 2) showed mean area-under-curve (AUC) of 0.70 based on T1w features. When MRF features were added to the T1w features, the mean AUC was 0.74. When FLAIR features were added to the T1w features, the mean AUC was 0.79. The best performance was shown by using all features including T1w, FLAIR and MRF, with mean AUC of 0.8. Figure 2 lower panels showcase two patients with FCD type II with accurate prediction results (Dice score > 90%).
Conclusions: Our study built a surface-based morphometric processing platform for automatic FCD detection using MRF data. The initial results revealed that implementing MRF as well as FLAIR features with T1w features improved vertex-level prediction performance. Individual-level performance optimization is ongoing.
Funding: NIH R01 NS109439, R21 EB026764
Neuro Imaging