The network is trained on 20 manually segmented thigh and leg datasets. It currently only works on out-phase gradient echo images, which can either be acquired or reconsturcted using dixon based methods. The implementation allows for any size data and works best is the data has a inplane resolution of 1-2.5 mm and a slice thinkness of 5-10 mm. The latest version of the trained networks are availible in the QMRItools package as .wlnet
files. Upon request *.ONNX
files are also availible. An example of a fully segmented leg dataset is shown below.
The segmentation networks are based on the UNET architectrue with ResNet convolution blocks that contain short skip connections. Below are the block diagrams of the encoding and decoding blocks used in our unet architecture.
For training the data is heavily augmented using: scale, skew, rotation, translation, noise, sharpen, contrast and and brightness. Training is done with a batch size of 2 and a patch size of 32x112x112 voxels and 256 datasets are seen per epoch. Patches are selected after data augmentation. Below are two examples of the first 100 epochs for the upper and lower leg. Training is typically continued for 200-300 epochs (8-12 hours). The resulting segmentation of both neworks is shown below.