The purpose of this project is to design, train and deploy a deep learning model on Azure Machine learning service which can diagnose a chest X-ray scan. The results of the diagnosis from the model can potentially serve as a preliminary diagnosis in a medical facility, resulting in reduced workload for radiologists and money and time saved for the medical facility.
A convenient webapp has been developed for easy and quick diagnosis of any chest X-ray scan. The source code for this is available here.
This project uses the NIH dataset for chest X-ray scans available here.
This project uses the NASnet architecture trained on ImageNet as a weights initializer, with a few layers added to adapt to the chest X-ray Dataset. The model have been trained using pretrained weights for transfer learning and also as a weights initializer. This model achieved 0.8182 AUC at 50 epochs compared to the original stanford paper which achieved 0.841 AUC.
| Model | AUC | Epochs |
|---|---|---|
| NASNet | 0.8182 | 50 |
| Disease | AUC Score | Disease | AUC Score |
|---|---|---|---|
| Atelectasis | 0.796015 | Pneumothorax | 0.866539 |
| Cardiomegaly | 0.866409 | Consolidation | 0.77846 |
| Effusion | 0.857553 | Edema | 0.837394 |
| Infiltration | 0.717564 | Emphysema | 0.915356 |
| Mass | 0.841312 | Fibrosis | 0.812732 |
| Nodule | 0.767547 | Pleural Thickening | 0.776329 |
| Pneumonia | 0.731321 | Hernia | 0.877378 |
If you just want to get the data, process it and train the model then run the scripts in the src/ folder in the following order after cloning this repo.
- get_and_extract_data.py
- data_prep.py
- pytorch_train.py The necessary packages are present in the env.yml file. It is recommended to use anaconda to setup the environment for you, using the env.yml file.
pytorch_run.py is a scoring script used with Azure machine learning service for the deployed model and the chest X-ray classfication web app.
- Clone this repo (for help see this tutorial).
- To utilize Azure Machine Learning Service run the 0_setup_aml notebook which will use your azure credentials and setup the Azure Machine learning workspace and storage account.
- Raw Data will be kept here within this repo. It can be downloaded by running src/get_and_extract.py or by running notebook/1_get_and_extract notebook
- For data processing and preparation, run 2_data_prep notebook or src/data_prep.py. This will process the dataset and store the train/validate/test split in data/processed/.
- (a) For training the model on Azure's compute target, run the 3_pytorch_train notebook notebook, this will setup the experiment and automatically procure a compute target and train the model there and store the output on the Azure workspace.
- (b) For training the model on your local system or a VM (e.g Azure Data Science Virtual Machine), simply run the following
python src/pytorch_train.py --data-dir="data/processed/images" - For Deploying the model on Azure, run the 4_pytorch_run noteboook. This will register the model,create the image and deploy on Azure Container Instances. The link to the deployment will be available on the azure portal.
This x-ray image pathology classification system is neither intended for real clinical diagnosis nor to be used as a substitute for a medical professional. The performance of this model for clinical use has not been established.
With deep gratitude to researchers and developers at PyTorch, NIH, Stanford, Project Jupyter and AzureChestXray repo. With special thanks to my mentors at Microsoft for their guidance.
This project welcomes contributions and suggestions. Please put up a Github issue before working on a bug or a feature, to avoid redundant work.
This project has adopted the Contributor Covenant Code of Conduct. For more information see Code of Conduct FAQ