Abstracts

Rationale: Despite advances in MRI analytics, current algorithms [1-4] do not detect >50% of focal cortical dysplasia (FCD) lesions [5]. Moreover, their deterministic nature [6] limits risk assessments of predictions, essential in clinical decision-making. Here, we propose an algorithm formulated on Bayesian convolutional neural networks (CNN) [7], providing prediction uncertainty. Our algorithm was trained and validated on data from 9 sites, for a total of 249 individuals. Methods: 3D T1-weighted and 3D FLAIR were acquired using the HARNESS-MRI [8] protocol in 148 patients with histologically-verified FCD (Table 1). The lesion was overlooked, and routine MRI deemed negative in 46%. This dataset served as inputs to a two-stage cascaded, uncertainty-informed Bayesian CNN classifier (Fig. 1): the first CNN optimized sensitivity by maximising the detection of lesional voxels; the second optimized specificity by reducing false positives. Sensitivity (prediction co-localizing with manual FCD labels) was tested using a leave-one-site-out strategy, whereby the training dataset comprised all sites except for the test site. Specificity was assessed by testing the model (trained on all sites) on 38 healthy controls (age: 30±7) and 63 disease controls (temporal lobe epilepsy patients, TLE; age: 31±8). Results: Overall sensitivity was 87% (129/148 FCD lesions detected; Fig. 2), with 6±5 extra-lesional clusters. Per-site sensitivity and false positive rates are summarized in Table 1. Notably, 76% of the lesions overlooked on routine MRI were detected by the classifier, and sensitivity was similar in MRI- and MRI+ cases. Specificity was 89% in healthy controls (4/38, 2±1 clusters) and 89% in TLE (7/63, 1±0). Conclusions: We present the first multicentre study with sites contributing data for both training and validation of an automated FCD detection algorithm based on deep learning. Our method, operating on multi-contrast MRI in voxel-space, demonstrated generalizability by showing robust performance across independent cohorts with varying age and scanner hardware. This deep learning approach is effective irrespective of lesion visibility on routine MRI. Moreover, our framework provides a degree of confidence through the inspection of the uncertainty map in individual patients (Fig. 2), thus allowing clinicians to assess the lesion in relation to putative false positives, possibly in conjunction with independent tests. This algorithm sets the basis for distributed machine learning across multiple sites through sharing of site-specific training models, rather than patient data, thus circumventing ethical challenges. Easy implementation, minimal pre-processing, and performance gains make this classifier the ideal platform for large-scale clinical use, particularly in MRI-negative FCD.References1. Adler S, et al. NeuroImage Clin. 20172. Hong S-J, et al. Neurology. 20143. Gill RS, et al. MICCAI. 20174. Tan Y-L, et al. Neuroimage. 2017 5. Kini LG, et al. NeuroImage: Clin. 2016 6. Gill RS, et al. MICCAI. 2018 7. Gal Y & Ghahramani Z. arXiv:stat.ML. 2015.8. Bernasconi A, et al. Epilepsia. 2019 Funding: CIHR (MOP-57840, MOP-123520)