This repo contains code for our NMR in Biomedicine paper - Deep learning methods for automatic segmentation of lower leg muscles and bones from MRI scans of children with and without cerebral palsy, which investigated and applied six deep learning models to segment individual lower leg muscles and bones from MRI scans of children with and without cerebral palsy.
The investigated models were: UNet, UNet++, 3D UNet, VNet, H-DenseUNet, and HybridUNet which is an in-house model inspired by H-DenseUNet. To learn more about UNet check out here. The paper where UNet++ was proposed can be found here, 3D UNet and VNet were proposed in 3D UNet and in VNet, Finally, H-DenseUNet was proposed in this paper.
To train a model, run the following command in the root directory (of this repo):
python3 ./entry.py --training_img_path [PATH_TO_YOUR_DATA]
--training_label_path [PATH_TO_YOUR_LABELS]
--model_name [NAME_OF_THE_MODEL] --val_img_path [PATH_TO_VAL_DATA]
--val_label_path [PATH_TO_VAL_LABELS]
--loss_fn [LOSS_FN_NAME] --optimiser [OPTIMISER_OF_CHOOSING] --num_epochs [NUM_EPOCHS
--output_path [YOUR_OUTPUT_PATH] --batch_size [BATCH_SIZE] --num_classes [NAME_OF_CLASSES] --learning_rate [LR_RATE]
--pre_training [TRUE/FALSE] --is_2d [TRUE/FALSE]Note that --is_2d needs to be True for 2D models and False for 3D models. Also note that the training data are expected to be NIFTI files. You need to modify data set classes if you wish to load data of other formats.