flkapes / deepray

DeepRay: Elevating X-Ray Diagnosis with AI-Powered Bone Image Classification

DeepRay

DeepRay is a collection of models and training code, built to classify X-Ray images of bones as normal or abnormal, such as with images in the MURA Dataset. This repository will contain the training code, and the code to convert trained models into an inference-compatible format for the framework BentoML. I have yet to publish the front end of this project, but I will soon. This was my dissertation project in my final year of University, and the goal here is to polish it up. Would love your thoughts on this, otherwise, enjoy!

Available Model Types

• DenseNet121
• DenseNet169
• DenseNet201
• ResNet101
• ResNet152
• ResNet101V2
• ResNet152V2
• InceptionV3
• InceptionResNetV2
• Xception

Dataset downloads & Pretrained Weights

Autodownload script, make sure that you have installed the requirements already:

  pip install -r requirements.txt
  python download.py --dataset
  python download.py --saved-checkpoints
  python download.py --mura-dataset

Download MURA directly from my Silicon Valley based Object Store

  wget -O MURA-v1.1.zip http://sjc1.vultrobjects.com/mura-dataset/MURA-v1.1.zip
  unzip MURA-v1.1.zip
  python extractimages.py
  rm MURA-v1.1.zip

Download the pre-prepared MURA Dataset Folder from my Silicon Valley based Object Store:

wget -O MURASeparated.zip http://sjc1.vultrobjects.com/mura-dataset/MURASeparated.zip
unzip MURASeparated.zip
rm MURASeparated.zip

Download Pretrained Weights Files from my Silicon Valley based Object Store

wget -O SavedCheckpoints.zip http://sjc1.vultrobjects.com/mura-dataset/SavedCheckpoints.zip
unzip SavedCheckpoints.zip
rm SavedCheckpoints.zip

Argument Descriptions

  -d <training_and_validation_path>                   -> /root/MURASeparated
  -l <learning_rate>                                  -> 7e-3
  -w <weight_decay>                                   -> 0
  -m <model_name>                                     -> densenet201
  -b <train_batch_size> <val_batch_size> <eval_size>  -> 48 24 0
  -s <training_validation_split>                      -> 0.2
  -e <max_epochs>                                     -> 48
  -p <body_part>                                      -> XR_FOREARM
  -H <image_input_size>                               -> 324
  -L <crop_image_size>                                -> 324
  -D <seed>                                           -> 44
  -T <path_to_weights_file>                           -> path/to/saved_checkpoints/
  -U <no_of_layers_to_finetune>                       -> 8
  --patience <early_stopping_patience>                -> 4

Run Locally

Go to the project directory, i.e. $HOME/Project/:

  cd Project/

Install dependencies:

  pip install -r requirements.txt

Enter src folder:

  cd src/

Run Training:

  python deepray.py -d <training_and_validation_path> -l 7e-3 -m resnet152 -b 48 24 0 -s 0.2 -e 48 -p XR_ELBOW -H 324 -L 324 -U 8 --patience 4

Run Evaluation / Load Weights:

  python deepray.py -d <evaluation_path> -l 7e-3 -m resnet152 -b 48 24 0 -s 0.2 -e 48 -p XR_ELBOW -H 324 -L 324 -U 8 --patience 4 -T /path/to/saved/weights/model.h5