medical-rib: A Python repository from Jameskry

Refed

Refed (Rib Extaction and Fracture Detection Model) by Youyou Jiang and Shiye Lei. Our model has a great performance on extracting ribs and detecting fracture from CT images. In Rib extraction, we can get 23.2 ribs (96.7%) in average for every patients; In Fraction Detection,recall: 71%, precision: 100%.

Refed is consist by two modules: rib extraction module and fracture detection module.

Workflow

read slices of CT data and reconstruct, see dcm read
separate bones using morphology and recognize all ribs, see separated, code see ribs_obtain
match labels to the ribs, see prepare_data, code see join
(Optional,only for train) data preparation for train data, voc2007, see prepare_data, code seevoc2007
yolo-v3 predict for demo/test or train for train data, see darknet/yolo-v3
(Optional,only for test)predict scores.

Installation and Dependencies

Install tensorflow. It is required that you have access to GPUs, The code is tested with Ubuntu 16.04 Python 3.6+, CUDA 9.0.176 and cudnn 7.4.2.
Clone the repository

    git clone https://github.com/jiangyy5318/medical-rib.git

Python dependencies (with pip3 install) or pip3 install -r requirements.txt:

    Deprecated==1.2.4
    image==1.5.27
    imageio==2.4.1
    interval==1.0.0
    lxml==4.2.5
    matplotlib==3.0.0
    numpy==1.15.2
    opencv-python==3.4.3.18
    pydicom==1.2.0
    pyparsing==2.2.2
    scikit-image==0.14.1
    scikit-learn==0.20.0
    scipy==1.1.0
    six==1.11.0
    nibabel==2.3.1
    pandas==0.23.4

Config darknet models

cd ${projects}/models

git clone https://github.com/pjreddie/darknet
cd darknet
vim Makefile

# update these three values.
GPU=1 # 0 if use cpu
CUDNN=1 # 0 if use cpu or cudnn not available
OPENCV=1 # 1 if use opencv else 0
NVCC=/path/to/nvcc

# build
make

cp darknet_cfg/cfg/* darknet/cfg/
cp darknet_cfg/data/hurt_voc.names darknet/data/

Demo and Test with pre-trained models

You can download pre-trained models, including GBDT model HERE, GBDT features HERE and yolo-v3 models HERE. Put feature.pkl, gbdt.pkl under the project root path (${project}/experiments/cfgs) and Put yolov3-voc_last.weights under the project root path (${projects}/models/darknet/backup)

    mkdir [demo_dir]
    mkdir [demo_dir]/pkl_cache/ # save desensitization data, pkl
    mkdir [demo_dir]/ribs_df_cache # save collected ribs (sparsely saved)
    mkdir [demo_dir]/voc_test_data # ribs projected on xoy 
    mkdir [demo_dir]/voc_test_predict # predict
     
    cd ${projects}
    # use raw data, slices of CT data
    ./experiments/scripts/demo.sh [DCM_PATH] [demo_dir]
    # use desensitization data, pkl data, CT data
    ./experiments/scripts/demo.sh [PKL_PATH] [demo_dir]
    ./experiments/scripts/demo.sh [demo_dir]/pkl_cache/patient_id.pkl [demo_dir]
    # DCM_PATH is folder path where CT slices existed.

Train your own model

Data Preparation

For dataSet structure and more information, follow the README under the preprocessing folder.

    ./experiments/scripts/nii_read.sh [DATA] [SLICING]
    ./experiments/scripts/dcm_read.sh [DATA]
    ./experiments/scripts/ribs_obtain.sh [LOGS_DIR] [FORMAT] [SLICING]
    ./experiments/scripts/prepare_data.sh
    # [DATA] in {updated48labeled_1.31, all_labeled} originated from different batches of data, has been defined in dcm_read.sh and nii_read.sh
    # [SLICING] in {0,1} is defined for interpolation interval, if 0, interval=1mm, else interval=slicing thickness 
    # [LOGS_DIR] used for debug
    # [FORMAT] '.dcm'

train gbdt model (generated `feature.pkl` and `gbdt.pkl`)

step ./experiments/scripts/ribs_obtain.sh will generate many separated bones and GBDT model will recognize all the ribs, all the features for every bone will be saved in the path ./data/bone_info_merges, you can added the misclassified bone in the path ./data/csv_files/update_err_bone_info.csv and run the below script, details see gbdt model

    ./experiments/scripts/generate_gbdt.sh

train your own yolo-v3 models

    cd ${projects}/models/darknet
    # only download once
    wget https://pjreddie.com/media/files/darknet53.conv.74
    ./darknet detector train ./cfg/hurt_voc.data ./cfg/yolov3-voc.cfg ./darknet53.conv.74 -gpus 0,1,2,3

to do list

refactoring

Jameskry/medical-rib