The available segmentation networks

With our framework we are able to automatically detect the body location. For each body part dedicated small light weight networks are developed and trained. If the data contains multiple locations the data is cut in various regions with considerable overlap and each is segmented by the correct neural network after which everything is collected in one segmentation file.

We aim to have segmentation networks for both GE and TSE Dixon based imaging that work on the raw magnitude data, in and out phase images and the reconstructed water and fat maps. Segmentation works best on 3D continuous datasets however those are not always available. Therefore we train our networks in 2D and 3D (except for the arm, see below). If data is acquired in 2D with slice gaps or extremely thick slices the 2D networks can be used. However 3D continuous data with 3D segmentation networks outperform the 2D variants consistently.

Networks included in QMRITools

Before segmentation, one classification network (Body_Pos_Side) determines the body location and side of the data. The following region networks are included.

The output of the networks is converted to standardized muscle labels. The full body label set (S1) is used by default, and older label sets for the legs and hip (S2), shoulder (S3) and upper arm (S4) are available. How to use the networks is described here, and how to train your own networks here.

• How to use the muscle segmentation framework
• Muscle segmentation naming
• The available segmentation networks
• Script based segmentation without mathematica
• Automated muscle segmentation

We gratefully acknowledge the support of NVIDIA with the donation of GPU used for this work.