Authors :
Ravnoor Gill, PhD – Montreal Neurological Institute & Hospital; Benoit Caldairou, PhD – Neuroimaging of Epilepsy Laboratory and Department of Neurology and Neurosurgery – Montreal Neurological Institute, McGill University, Montreal, Canada; Neda Bernasconi, MD, PhD – Neuroimaging of Epilepsy Laboratory and Department of Neurology and Neurosurgery – Montreal Neurological Institute, McGill University, Montreal, Canada; Andrea Bernasconi, MD – Neuroimaging of Epilepsy Laboratory and Department of Neurology and Neurosurgery – Montreal Neurological Institute, McGill University, Montreal, Canada
This abstract has been invited to present during the Neuroimaging platform sessionRationale: Personalized diagnostics in temporal lobe epilepsy (TLE) motivate accurate automated segmentation of the hippocampus, the hallmark site of pathology. We combined a volume-based subfield segmentation method, DeepPatch, relying on convolutional neural networks (CNN) with a linear discriminant analysis classifier (LDA) to lateralize the seizure focus, both in patients with obvious hippocampal sclerosis (HS) and those considered to have MRI-negative HS based on visual evaluation.
Methods: Training. We trained our algorithm (DeepPatch) on freely-available hippocampal subfield labels segmented manually on 0.6 mm isotropic T1-weighted MRI of 25 healthy subjects (31±7 yrs, 13 females) [1]. For cross-validation, the dataset was partitioned to allocate 37.5% hippocampi for training, 37.5% for the atlas (to sample similar patches), and 25% for testing. A similarity function [2] matched the intensity of randomly selected patches of the training set to the atlas. All patches were used to train a CNN to model multiscale intensity features and implicitly learn between-subfields boundaries.
Testing. For each hippocampus, we extracted a patch around each voxel, matched it to the most similar atlas patches, and fed them to the CNN. Resulting probabilistic labels were aggregated through majority voting to produce the final segmentation.
Validation. A clinical dataset of 76 TLE patients with histologically confirmed HS (35±10 yrs, 47 females; 39/76 MRI-negative) were used for validation.
Performance. Dice index and Bland-Altman plots evaluated overlap and compared automated to manual labels, respectively. The LDA classifier using volumes and intensity derived from T2-weighted and FLAIR/T1 contrasts was used to lateralize the focus [3].
Results: In healthy controls, the
average overlap between manual and automated labels was 91.9%±1.2 for CA1-3, 87.9%±2.2 for CA4-DG and 88.9%±2.0 for subiculum; similar performance was obtained in TLE, with a Dice of 90.7%±2.3, 86.6%±4.5 and 87.8%±2.5. High correlations and small differences between in volume between automated and manual segmentations further supported robustness (Figure 1); examples are shown in Figure 2. Higher accuracy for seizure focus lateralization was obtained using T2 compared to FLAIR/T1 (90.8±2.2% vs. 85.7±2.1%; p< 0.05); both contrasts were superior to volumetry (69.7±3.2%). The combination of T2 and FLAIR/T1 yielded the best performance, with 91.4±2.7% in MRI-negative and 99.5±1.1% in MRI-positive patients.