RadNet
Package for bio-medical image segmentation.
Getting Started • Train • Test • Interpret • Performance • Release Notes • Upcoming Releases • Citation • FAQ • Blog
Made by Mukesh Mithrakumar • 🌌 https://mukeshmithrakumar.com
What is it ❓
**RadNet** is an ensemble convolutional neural network package (using U-Net, VGG, and Resnet) for biomedical image
detection, segmentation and classification.
Currently the code works for the ISBI Neuronal Stack Segmentation dataset. See Release Notes for the current release features and see Upcoming Releases for the next release enhancements.
If this repository helps you in anyway, show your love
Please Note that since this is a developer release content is being constantly developed and till I test everything completely I won't be committing updates into the repo so if you run into any issues, please reach out. The best way to prevent this is to use the released source for developing.
📋 Getting Started
📀 Software Prerequisites:
To see the software prerequisites (click to expand...)
```
- pip install 'matplotlib'
- pip install 'graphviz'
- pip install 'tensorflow'
- pip install 'scikit-learn'
- pip install 'tifffile'
- pip install 'Pillow'
- pip install 'scipy'
- pip install 'numpy'
- pip install 'opencv-python>=3.3.0'
- pip install 'torch'
- pip install 'torchvision'
- pip install 'pytest'
- pip install 'flake8'
- pip install 'cython'
- pip install 'psutil'
```
💻 Hardware Prerequisites:
Runs on a NVIDIA GeForce GTX 1050 Ti with 4 GB GDDR5 Frame Buffer and 768 NVIDIA CUDA® Cores.
📘 Folder Structure
To see the folder structure (click to expand...)
```
main_dir
- data (The folder containing data files for training and testing)
- pytorch_unet (Package directory)
- model (PyTorch u-net model)
- u_net.py
- optimize
- c_extensions.pyx
- config.py
- hyperparameter.py
- multi_process.py
- performance.py
- processing
- augments.py
- load.py
- trainer
- evaluate.py
- interpret.py
- train.py
- utils
- helpers.py
- metrics.py
- unit_test.py
- visualize
- logger.py
- plot.py
- train_logs (will be created)
- visualize (will be created)
- weights (will be created)
```
🔧 Install
Currently you can clone the repo and start building, mean while, am working on the PyPi release, so will be updated
⌛ Train
Train the model by running:
train.py root_dir(path/to/root directory)
Arguments that can be specified in the training mode:
usage: train.py [-h] [--main_dir MAIN_DIR] [--resume] [-v]
[--weights_dir WEIGHTS_DIR] [--log_dir LOG_DIR]
[--image_size IMAGE_SIZE] [--batch_size BATCH_SIZE]
[-e EPOCHS] [-d DEPTH] [--n_classes N_CLASSES]
[--up_mode {upconv, upsample}] [--augment]
[--augment_type {geometric, image, both}]
[--transform_prob TRANSFORM_PROB] [--test_size TEST_SIZE]
[--log] [-bg]
Script for training the model
optional arguments:
-h, --help show this help message and exit
--main_dir MAIN_DIR main directory
--resume Choose to start training from checkpoint
-v, --verbose Choose to set verbose to False
--weights_dir WEIGHTS_DIR
Choose directory to save weights model
--log_dir LOG_DIR Choose directory to save the logs
--image_size IMAGE_SIZE
resize image size
--batch_size BATCH_SIZE
batch size
-e EPOCHS, --epochs EPOCHS
Number of training epochs
-d DEPTH, --depth DEPTH
Number of downsampling/upsampling blocks
--n_classes N_CLASSES
Number of classes in the dataset
--up_mode {upconv, upsample}
Type of upsampling
--augment Whether to augment the train images or not
--augment_type {geometric, image, both}
Which type of augmentation to choose from: geometric,
brightness or both
--transform_prob TRANSFORM_PROB
Probability of images to augment when calling
augmentations
--test_size TEST_SIZE
Validation size to split the data, should be in
between 0.0 to 1.0
--log Log the Values
-bg, --build_graph Build the model graph
📋 Logging
To activate logging of the errors (:default is set as no)
train.py root_dir(path/to/root directory) --log
To see the log in tensorboard follow the log statement after training:
📊 Network Graph
Since Pytorch graphs are dynamic I couldn't yet integrate it with tensorflow but as a quick hack run the following to build a png version of the model architecture (:default is set as no)
train.py root_dir(path/to/root directory) -bg
To see the output of the graph (click to expand...)
⌚ Test
Evaluate the model on the test data by running:
evaluate.py root_dir(path/to/root directory)
Arguments that can be specified in the evaluation mode:
usage: evaluate.py [-h] [--main_dir MAIN_DIR] [--image_size IMAGE_SIZE]
[--weights_dir WEIGHTS_DIR]
Script for evaluating the trained model
optional arguments:
-h, --help show this help message and exit
--main_dir MAIN_DIR main directory
--image_size IMAGE_SIZE
resize image size to match train image size
--weights_dir WEIGHTS_DIR
Choose directory to save weights model
📉 Interpret
Visualize the intermediate layers by running:
interpret.py root_dir(path/to/root directory)
Arguments that can be specified in the interpret mode:
usage: interpret.py [-h] [--main_dir MAIN_DIR]
[--interpret_path INTERPRET_PATH]
[--weights_dir WEIGHTS_DIR] [--image_size IMAGE_SIZE]
[--depth DEPTH]
[--plot_interpret {sensitivity,block_filters}]
[--plot_size PLOT_SIZE]
Script for interpreting the trained model results
optional arguments:
-h, --help show this help message and exit
--main_dir MAIN_DIR main directory
--interpret_path INTERPRET_PATH
Choose directory to save layer visualizations
--weights_dir WEIGHTS_DIR
Choose directory to load weights from
--image_size IMAGE_SIZE
resize image size
--depth DEPTH Number of downsampling/upsampling blocks
--plot_interpret {sensitivity,block_filters}
Type of interpret to plot
--plot_size PLOT_SIZE
Image size of sensitivity analysis
🔩 Sensitivity Analysis
To do sensitivity analysis run:
interpret.py root_dir(path/to/root directory) --plot_interpret sensitivity
🔩 Block Analysis
To visualize the weight output of each up/down sampling block run:
interpret.py root_dir(path/to/root directory) --plot_interpret block_filters
📈 Performance
(Work in Progress)
Release Notes
- Works for the ISBI Neuronal Stack Segmentation dataset
- Tensorboard logging to visualize training/validations/dice scores
- Build for PyTorch static graph
- Intermediate filter layer visualizations
Upcoming Releases
Keep an eye out 
- performance.py to measure code performance and optimize code
- hyperparameters.py to tune hyper parameters
- config.py with configs for the hyper parameters
- multi gpu/multi processing data training.
- c extensions for certain functions
- complete unit_test.py for the above
- code coverage to check tests and iterate
- Pre alpha PyPI package
- Biomedical image pre-processing script
- modifications for the unet to work on MRI data
- test on the CHAOS Segmentation challenge
- modifications for the unet to work on CT scan
- test on the PAVES Segmentation challenge
- complete unit_test.py for the above
- Deploy alpha PyPI package
- Neural architecture search script
- Classifier to identify between the organs (One U-Net to segment different organs)
- Separate classifier to identify different cells
- Deploy alpha PyPI package
- Graphical user interface for RadNet
- Developer and researcher mode for the GUI
- Abstracted away the deep learning stuff so its not python/deep learning friendly but more like doctor friendly
- Build into a software package
- Deploy beta PyPI package
©️ Citation
💬 FAQ
- For any questions and collaborations you can reach me via LinkedIn