spratyush02/RSNA-Intracranial-Hemorrhage-Detection-using-Deep-Learning

We present a method to correctly predict presence of Intracranial Hemorrhage and identify its type. Our model imitates the procedure followed by radiologists to analyse a 3D CT scan in real-world. The model utilizes multi-window 3D context from neighboring slices to improve predictions at each slice and subsequently, aggregates the slicelevel predictions to provide patient diagnosis for Intracranial Hemorrhage . Our proposed architecture performs significantly better than standard single window based non-contextual models

RSNA Intracranial Hemorrhage Detection Instructions

Please find the project report in this repository

Table of contents

Setup
File Structure
Files to be downloaded
Data preprocessing and preparation
Extracting metadata from DICOM files
Processing Train Labels
Converting raw images to tfrecords format
Training the Base model
Extract base model embeddings for Sequence model
Training Sequence (LSTM) model
Prediction using saved checkpoint

Setup

Required Python >= 3.6.0

Create virtual environment

python -m venv .venv

Activate environment (Windows)

.venv\Scripts\activate.bat

Activate environment (Unix or MacOS)

source .venv/bin/activate

Installing dependencies (for CPU)

pip install -r requirements.cpu.txt

Installing dependencies (for GPU, recommended)

pip install -r requirements.gpu.txt

File Structure

Save all the files in same folder except wherever mentioned. Also do not change the names of downloaded files

Following should the file structure to run the model smoothly

data/
    intracranial-hemorrhage-detection.zip (181G)
    rsna-intracranial-hemorrhage-detection/ (~13G)
        stage_2_train/
        stage_2_test/
    test/
        test.tfrecord (1.8G)
    train/
        train-01-of-20.tfrecord (545M)
        train-01-of-20.tfrecord
        .....
    validation/
        train-19-of-20.tfrecord (545M)
        train-20-of-20.tfrecord
    train_metadata.csv (80M)
    test_metadata.csv (13M)
    stage_2_train.csv (120M)
    processed_stage_2_train.csv (263M)
    reshaped_stage_2_train.csv (19M)
    train_embeddings.npy (14G)
    train_labels.npy (27M)
    train_mask.npy (9M)
    test_embeddings.npy (2.5G)
    test_img_ids.npy (9.7M)
experiments/
    1577737599_tf_bin_loss_b32_e10/
    lstm/
        1578060580_b32_e20_l3_cs128/
src/
    create_metadata.py
    .....
inception_resnet_v2_2016_08_30.ckpt (237M)
resnet_v2_50.ckpt (307M)

Files to be downloaded

• Exracted Images
• Metadata
• Train Labels
• Processed Labels
• TFrecord Files
• Pretrained Checkpoints
• Base Model Checkpoint
• Sequence Model Data
• Sequence Model Checkpoint

Data preprocessing and preparation

Requires 5G memory
To extract train and test images from DICOM files, download the raw rsna-intracranial-hemorrhage-detection.zip file here. After this run the following command

python src/img_extract.py --zip_dir <zip file directory path>

• zip_dir: Directory path where zip file is present. Please note that all the future data will be stored in this directory

After this train images will be stored in zip_dir/rsna-intracranial-hemorrhage-detection/stage_2_train folder and test images will be stored in zip_dir/rsna-intracranial-hemorrhage-detection/stage_2_test folder. Since this process takes some so extracted images can be downloaded here and save it in zip_dir

Extracting metadata from DICOM files

Requires 5G memory
To extract train and test metadata, make sure that you have downloaded raw .zip file. Then run the following command

python src/create_metadata.py --zip_dir <zip file directory path>

• zip_dir: Directory path where zip file is present. Please note that all the future data will be stored in this directory

After this train_metadata.csv and test_metadata.csv will be created in zip_dir. These files can be downloaded here and save it in zip_dir

Processing Train Labels

To process train labels, so that it can be fed to tfrecords creation step, download the train labels files here and save it in same folder where raw .zip file is saved (zip_dir). The run the following command

python src/process_labels.py --data_dir <train label csv file directory>

• data_dir: Directory where train label csv file is saved

After this processed_stage_2_train.csv and reshaped_stage_2_train.csv will be created in data_dir. These files can be downloaded here and save it in data_dir

Converting raw images to tfrecords format

To convert raw images and labels to tfrecords files, make sure you have extracted images from dicom files and have also processed the train labels csv file (processed_stage_2_train.csv) as explained in previous step in same data directory. After all the raw data is prepared, run the following command

python src/create_tf_records.py --data_dir <directory path where all the data is being stored>

• data_dir: Directory path where all the data including tfrecords files will be saved

After this three folder naming train, validation and test will be created in data_dir and you can find tfrecord files there. Since this process takes some time, so you can download already created tfrecord files here and save it in data_dir

Training the Base model

Requires 10G memory
To train the model you need to have saved checkpoint for resnet_v2 and inception_resnet. Both checkpoints which can be downloaded from here. Download both the checkpoints in current working directory. Then run following command

python src/train.py [--batch_size] [--num_epochs] [--lr] [--pretrained_model] [--save_dir] [--data_dir] [--loss] [--export]

• batch_size: default values is 32
• num_epochs: default values is 5
• lr: default value is 0.0001. Learning rate for training
• pretrained_model: default value is inception_resnet. Other possible value is resnet_50
• save_dir: default value is "./experiments". Directory where model and config files are saved
• data_dir: Directory path where all the data including tfrecords files are saved
• loss: default value is "bin_loss". Loss to be used for training
• export: Save model checkpoint or not?

After running this model checkpoints, config file and tensorboard files will be stored in directory naming save_dir/<timestamp>_tf_<loss_name>_b<batch_size>_e<epoch_num>

To launch the tensorboard file while training run the following command

tensorboard --logdir current/working/directory/save_dir/<timestamp>_tf_<loss_name>_b<batch_size>_e<epoch_num>

Saved checkpoints for base model training can be downloaded here

Extract base model embeddings for Sequence model

Requires 70G memory
To extract base model embeddings which can then be used for sequence model, make sure you have base model checkpoints saved in current/working/directory/experiments/<timestamp>_*. Also tfrecord files, metadata csv files and reshaped labels csv file (reshaped_stage_2_train.csv) is present in data_dir. Then run the following command

python src/extract_features.py --model_id <saved model id> [--base_data] [--max_to_keep] --data_dir <saved data directory path>

• model_id: model_id (timestamp) of saved model
• base_data: Whether to extract train data embeddings or test data embeddings. Default value is train. Possible values are train and test
• max_to_keep: How many number of timestamps data is needed in unrolled sequence model. Default is 60
• data_dir: Directory where tfrecord files, metadata csv files and reshaped labels csv file (reshaped_stage_2_train.csv) is saved

After this different files will be stored in data directory based on whether base_data is train or test.

• If base_data is train, then three .npy files will be created naming:
  • train_embeddings.npy: Embedding tensor of shape (patient_count, max_to_keep, base_model_final_layer_size). base_model_final_layer_size = 1536 for inception resnet and 2048 for resnet50 model.
  • train_labels.npy: Label tensor of shape (patient_count, max_to_keep, 6)
  • train_mask.npy: Mask tensor of shape (patient_count, max_to_keep). It has binary values having zero where padding is used else one. If base_data is test, then two .npy files will be created naming:
  • test_embeddings.npy: Same as train_embedding.npy
  • test_img_ids.npy: Tensor of image ids of shape (patient_count, max_to_keep). It will later be used to map final predictions with image ids

Since this step can take some time, so all the required npy files can be downloaded here and save it in data_dir

Training Sequence (LSTM) model

Requires 70G memory
To train Sequence (LSTM) model make sure train_embeddings.npy, train_labels.npy and train_mask.npy is saved in data_dir. Then run the following command:

python src/train_lstm.npy [--batch_size] [--num_epochs] [--num_layers] [--cell_size] [val_ratio] [--lr] [--dropout_rate] [--save_dir] [--data_dir] [--loss] [--export]]

• batch_size: default values is 32
• num_epochs: default values is 10
• num_layers: default value is 2
• cell_size: default value is 256
• val_ratio: default value is 0.1
• lr: default value is 0.0001. Learning rate for training
• dropout_rate: default value is 0.7
• save_dir: default value is "./experiments/lstm". Directory where model and config files are saved
• data_dir: Directory path where all the data including npy are saved
• loss: default value is "masked_bin_loss". Loss to be used for training
• export: Save model checkpoint or not?

After running this model checkpoints, config file and tensorboard files will be stored in directory naming save_dir/<timestamp>_b<batch_size>_e<epoch_num>_l<num_layers>_cs<cell_size>

To launch the tensorboard file while training run the following command

tensorboard --logdir current/working/directory/save_dir/<timestamp>_b<batch_size>_e<epoch_num>_l<num_layers>_cs<cell_size>

Saved checkpoints for seqence model training can be downloaded here. Save the checkpoints in experiments/lstm/<timestamp>_*

Prediction using saved checkpoint

Requires 10G memory
To predict on test set make sure that you have saved checkpoint files, test.tfrecord, test_embeddings.npy and test_img_ids.npy in correct folders. Prediction can be made on base model as well as sequence model. Run the following command

python src/predict.py --model_id <saved model id> [--model_type] --data_dir <saved data directory path>

• model_id: model_id (timestamp) of saved model
• model_type: Whether to make prediction on base model or lstm model. Default value is lstm. Possible values are lstm and base
• data_dir: Directory path where all the data including npy are saved

After running this prediction file on test data will be generated in current working directory named output2.csv

Write about memory requirements and time overall bsub command. Also remove option of save dir