AnatomyAware Nets v2
Currently just a collection of three segmentation networks (U-Net [1], PSP-Net [2] (with resnet extractors [3]), DeepMedic [4]), that allow for three different inclusions of wider context (skip connections, pyramid pooling - dilated convolutions, downsampled pathway), useful particularly when training on smaller patches.
To run the training with any of these networks on your own data, simply comment out sys.argvs line in Training.py (useful for debugging), and run
python3 Training.py --flags
in the terminal, together with desired flags and parameters (see argument parser in Training.py for more info). After training is finished, the results will be saved in folder RESULTS: a file with trained pytorch network checkpoint, a txt file with all set parameters and a csv file with all metrics info per epoch.
Your data should be structured in the following way:
├── ...
├── Training.py # Main script for training
├── DATAFOLDER # Main Data folder. Make one for each individual dataset.
│ ├── TRAIN # Folder for training data
│ | ├── in1 # main input images/patches, in npy format
│ | ├── in2 # corresponding downsampled input images/patches, in npy format (only used for DeepMedic)
│ | └── gt # correspoinding ground truth images/patches
│ ├── VAL # Folder for validation data
│ ├── in1 # main input images/patches, in npy format
│ ├── in2 # corresponding downsampled input images/patches, in npy format (only used for DeepMedic)
│ └── gt # correspoinding ground truth images/patches
└── ...
The script Slicing.py contains (hardcoded to my own dataset) code for slicing patches and saving them as .npy under an appropriate structure. The script Postprocessing.py contains (again somewhat hardcoded to my own data) functions for ad-hoc visualizing and comparing training curves and segmentation output.
[1] U-Net paper: https://arxiv.org/pdf/1505.04597.pdf
[2] PSP-Net paper: https://arxiv.org/pdf/1612.01105.pdf
[3] extractors implementation shamelesly taken from: https://github.com/Lextal/pspnet-pytorch/blob/master/extractors.py
[4] DeepMedic paper: https://www.sciencedirect.com/science/article/pii/S1361841516301839