CAN3D: FAST 3D KNEE MRI SEGMENTATION VIA COMPACT CONTEXT AGGREGATION

Automated segmentation using deep learning approaches have shown significant promise for medical images. However, existing methods generally suffer from high computational complexity when utilised in 3D due to their large memory requirements, thus restricting training to high-performing computing hardware only. We present an extremely compact convolutional neural network with a shallow memory footprint to address this problem and train the model with a novel loss function to segment imbalanced classes with extra shape constrain in 3D MR images. The proposed approaches can directly process large full-size 3D input volumes (no patches) and allow inference times within just seconds using the CPU. The proposed network efficiently retains model parameters required to outperform other methods for 3D segmentation (U-Net3D, improved U-Net3D and V-Net) under severe memory limitations, while achieving several times faster inference times. It can also achieve favourable performance compared to a complex pipeline for knee cartilage segmentation with hundred times faster inference.