Regularized siamese neural network for unsupervised outlier detection on brain multiparametric magnetic resonance imaging: Application to epilepsy lesion screening

• Published in: Medical image analysis Vol. 60; p. 101618
• Main Authors: Alaverdyan, Zaruhi, Jung, Julien, Bouet, Romain, Lartizien, Carole
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.02.2020; Elsevier BV; Elsevier
• Subjects: Adolescent; Adult; Aged; Anomalies; Anomaly detection; Brain; Brain lesions; Brain mapping; Computer Science; Data analysis; Datasets as Topic; Deep learning; Epilepsy; Epilepsy - diagnostic imaging; Feature extraction; Female; Heterogeneity; Humans; Image Interpretation, Computer-Assisted - methods; Lesions; Localization; Machine Learning; Magnetic Resonance Imaging; Male; Medical Imaging; Middle Aged; Neural networks; Neural Networks, Computer; Neuroimaging; Outliers (statistics); Regularization; Regularized siamese network; Representations; Signal and Image Processing; Spatial discrimination; Unsupervised Machine Learning; Unsupervised representation learning; Wasserstein autoencoder; Wasserstein autoencoder; Deep learning; Brain lesions; Anomaly detection; Regularized siamese network; Unsupervised representation learning; Regularized Siamese network; Deep Learning
• ISSN: 1361-8415; 1361-8423; 1361-8423
• DOI: 10.1016/j.media.2019.101618

•We cast the challenging problem of detecting subtle brain lesions as a per voxel outlier detection problem.•Our brain anomaly detection model is trained on normal subjects only.•Our model combines unsupervised latent representation with a novel deep siamese network and one-class classification.•Patients with intractable epilepsy are considered normal on MRI (MRI negative) in 30–80% cases.•Our model detects 61% of MRI-negative epilepsy lesions in multi-parametric MRI (T1/FLAIR) while human performance is at 0%. [Display omitted] In this study, we propose a novel anomaly detection model targeting subtle brain lesions in multiparametric MRI. To compensate for the lack of annotated data adequately sampling the heterogeneity of such pathologies, we cast this problem as an outlier detection problem and introduce a novel configuration of unsupervised deep siamese networks to learn normal brain representations using a series of non-pathological brain scans. The proposed siamese network, composed of stacked convolutional autoencoders as subnetworks is designed to map patches extracted from healthy control scans only and centered at the same spatial localization to ‘close’ representations with respect to the chosen metric in a latent space. It is based on a novel loss function combining a similarity term and a regularization term compensating for the lack of dissimilar pairs. These latent representations are then fed into oc-SVM models at voxel-level to produce anomaly score maps. We evaluate the performance of our brain anomaly detection model to detect subtle epilepsy lesions in multiparametric (T1-weighted, FLAIR) MRI exams considered as normal (MRI-negative). Our detection model trained on 75 healthy subjects and validated on 21 epilepsy patients (with 18 MRI-negatives) achieves a maximum sensitivity of 61% on the MRI-negative lesions, identified among the 5 most suspicious detections on average. It is shown to outperform detection models based on the same architecture but with stacked convolutional or Wasserstein autoencoders as unsupervised feature extraction mechanisms.