Abstracts

Rationale: Focal cortical dysplasia (FCD) is a common cause of pharmacoresistant epilepsy, often requiring surgical resection. Detecting FCD on MRI can be challenging and labor intensive. Deep-learning models can identify high-level patterns indicative of FCD (Gill et al, Neurology 2021). No previous studies in FCD detection have combined quantitative MRI with deep learning. Magnetic Resonance Fingerprinting (MRF) is a novel quantitative imaging technique that provides fast, reliable tissue property measurements (Ma et al, Nature 2013). In this study, we trained and validated a deep-learning pipeline using MRF and clinical MRI data for whole-brain FCD detection.

Methods: 3D whole-brain MRF and T1w MPRAGE scans were acquired from 67 healthy controls (HCs) and 40 patients with FCD (11 Type IIA, 13 Type IIB, 13 mMCD, and 2 MOGHE; 36 pathologically confirmed) on a 3T SIEMENS Prisma scanner. MRF T1, T2, GM and WM maps, 3D lesion labels were generated as previous described (Ding et al, Epilepsia 2024). All images were registered into the MNI space. Each subject’s T1 and T2 maps were smoothed using a 3×3×3 averaging kernel excluding CSF voxels. Mean and standard deviation T1 and T2 maps (mT1, mT2, sdT1, and sdT2) were then calculated voxel-wise from the smoothed maps across all HCs. T1 and T2 z-scores maps were calculated by subtraction of the mean map and division by the sd map. MRF junction and extension z-score maps were calculated in the same manner as MAP (Huppertz et al, Epilepsia 2008), using the MRF GM and WM probability maps as inputs. No-new U-Net (nnUNet) was used for the deep-learning model (Isensee et al, Nature Methods 2021). We used T1w MPRAGE, synthetic MRF T1w image, MRF T1 and T2, MRF T1 and T2 z-score maps, and MRF tissue probability maps as potential multiparametric inputs to nnUNet and trained for 250 epochs with leave-one-patient-out cross validation. Model parameters were selected from the epoch with the best pseudo-dice score. Additional post-processing was applied to reduce false positive (FP) clusters. Model performance at the patient level was evaluated. The workflow is shown in Figure 1.

Results: Using the T1w MPRAGE, MRF z-score maps, and MRF tissue probability maps as model inputs resulted in the best performance. The patient level-sensitivity was 85% and 4.7 FPs per patient (FP/pt). Further location probability-based processing reduced the FPs to 2.8 FP/pt while maintaining 80% sensitivity. Figure 2 shows examples of detected lesion clusters (white) which overlapped well with the lesion labels (cyan outline). Using the T1w image and MRF z-score maps only (without tissue probabilities) resulted in 70% sensitivity and 5 FP/pt. Using only the T1w MPRAGE resulted in less than 70% sensitivity when trained on patients only and on all subjects. Training on the synthetic MRF T1w image with or without raw MRF T1 and T2 data across all subjects resulted in less 60% sensitivity but less than 2 FP/pt.


Conclusions: We show the initial efficacy of a deep-learning pipeline for MRF-based whole-brain FCD detection. Normalized MRF maps and tissue probabilities are valuable features that can improve the performance of deep-learning models in addition to T1w input.

Funding: NIH R01 NS109439