A Deep Learning Model based on MobileNetV3 and UNet for Spinal Cord Gray Matter Segmentation

Changes in the gray matter (GM) tissue of the human spinal cord may indicate a wide range of neurological disorders. Thus, the detection and segmentation of GM regions in Magnetic Resonance Imaging (MRI) is an important task when studying the spinal cord and its related medical conditions. In this work, we propose a new method for the segmentation of GM tissue in spinal cord MRI images based on deep convolutional neural networks. Our proposed method, called MobileNet-V3-UNet, uses the recent light-weight pre-trained MobileNet-V3 CNN model (large version) as a backbone for feature extraction, augmented with a set of up-sampling layers and skip connections similar to the UNet architecture. We explain in our paper how the proposed new architecture is built and trained, then we test it on the spinal cord GM challenge dataset. The obtained preliminary results show some good capabilities of the proposed approach, as it has outperformed three previous methods with respect to several evaluation metrics. Another advantage of our method is the low computational requirements as the number of trainable parameters in the proposed model is 7,843,587 only.