A Structured Deep-learning Based Approach for the Automated Segmentation of Human Leg Muscle from 3D MRI

In this paper, we present an automated algorithm for segmenting human leg muscles from 3D MRI data using deep convolutional neural network (CNN). Using a generalized cylinder model the human leg muscle can be represented by two smooth 2D parametric images representing the contour of the muscle in the MRI image. The proposed CNN algorithm can predict these two parametrized images from raw 3D voxels. We use a pre-trained AlexNet as our baseline and further fine-tune the network that is suitable for this problem. In this scheme, AlexNet predicts a compressed vector obtained by applying principal component analysis, which is then back-projected into two parametric 2D images representing the leg muscle contours. We show that the proposed CNN with a structured regression model can out-perform conventional model-based segmentation approach such as the Active Appearance Model (AAM). The average Dice score between the ground truth segmentation and the obtained segmentation image is 0.87 using the proposed CNN model, whereas for AAM score is 0.68. One of the greatest advantages of our proposed method is that no initialization is needed to predict the segmentation contour, unlike AAM.