Convolutional neural networks to identify malformations of cortical development: A feasibility study.

Objective: To develop and test a deep learning model to automatically detect malformations of cortical development (MCD).

Methods: We trained a deep learning model to distinguish between diffuse cortical malformation (CM), periventricular nodular heterotopia (PVNH), and normal magnetic resonance imaging (MRI). We trained 4 different convolutional neural network (CNN) architectures. We used batch normalization, global average pooling, dropout layers, transfer learning, and data augmentation to minimize overfitting.

Results: There were 45 subjects (866 images) with a normal MRI, 52 subjects (790 images) with CM, and 32 subjects (750 images) with PVNH. There was no subject overlap between the training, validation, and test sets. The InceptionResNetV2 architecture performed best in the validation set in all models and was evaluated in the test set with the following results: 1) the model distinguishing between CM and normal MRI yielded an area under the curve (AUC) of 0.89 and accuracy of 0.81; 2) the model distinguishing between PVNH and normal MRI yielded an AUC of 0.90 and accuracy of 0.84; 3) the model distinguishing between the three classes (CM, PVNH, and normal MRI) yielded an AUC of 0.88 and accuracy of 0.74. Visualization with gradient-weighted class activation maps and saliency maps showed that the deep learning models classified images based on relevant areas within each image.

Significance: This study showed that CNNs can detect MCD at a clinically useful performance level with a fully automated workflow without image feature selection.