A dual encoder–decoder multi-task 3D deep learning framework for the segmentation of focal cortical dysplasia lesions

Focal cortical dysplasia (FCD) is a congenital malformation of brain development that is the most common cause of intractable epilepsy in adults and children. Automating the identification and segmentation of FCD lesions from magnetic resonance imaging (MRI) volumes is useful for neuroradiologists in pre-surgical evaluations. The prevailing methods in FCD segmentation using two-dimensional (2D) convolutional neural networks (CNNs) largely overlook the potential of utilizing three-dimensional (3D) MRI volumes, thus neglecting the valuable inter-slice information inherent in the MRI volumes. We propose a novel 3D deep learning model employing a multi-view dual encoder–decoder architecture to precisely segment FCD lesions within MRI volumes. Our approach is based on a 3D CNN framework with integrated residual connections, serving as the backbone for the segmentation network. The model also incorporates various architecture-wise enhancements. Firstly, we integrate multi-view training, a concept drawn from the methodology employed by neuro-radiologists when examining 3D MRI volumes. Here, the model processes fluid-attenuated inversion recovery (FLAIR) MRI volumes and their corresponding cortical thickness maps. This information is channeled through a dual-encoder network, wherein the individual encoders are interlinked through a 3D attention mechanism. Additionally, the model implements a dual-decoder stage to facilitate dual-task learning, leveraging the distance map derived from the ground truth data. The model achieved Dice similarity coefficient (DSC) that were 4.8% and 2.3% higher compared to state-of-the-art 2D and 3D FCD segmentation methods, respectively.