DeepLesion

This repository converts the DeepLesion dataset for use with mmdetection.

DeepLesion is a large-scale dataset for medical object detection. This dataset contains over 32,000 annotated CT images and is split into train, validation, and test sets. It coarsely defines eight types of lesions, including bone, abdomen, mediastinum, liver, lung, kidney, soft tissue, and pelvis. We have treated all lesions under the same 'lesion' category.

1️⃣ Preparing the Images

git clone https://github.com/anir16293/Deep-Lesion.git

Modify the download_directory in batch_download_zips.py

# it takes a while, so using tmux is recommended
python batch_download_zips.py

In unzip_me.py, set base_dir to the directory of the downloaded zip files and set a new output_dir

# it takes a while, so using tmux is recommended
python unzip_me.py

If cv2 is not installed, run pip install opencv-python to install it

Modify the path in preprocess_me.py. Set csv_path to the location of DL_info.csv

python preprocess_me.py

2️⃣ Preparing the Annotations

In coconize_me.py, insert the location of DL_info.csv into csv_file_path

python coconize_me.py

3️⃣ Inject noise using UNA (optional)

git clone https://github.com/Ryoo72/UNA.git
cd UNA
python una_inj.py --ratio 0.1 --class_type deeplesion --path {path_to_json} --target {target_path}
# EXAMPLE
# python una_inj.py --ratio 0.1 --class_type deeplesion --path ./annotations/deeplesion_train.json --target ./annotations

4️⃣ Setting Up MMDetection

git clone https://github.com/open-mmlab/mmdetection.git

Duplicate configs/_base_/datasets/coco_detection.py → configs/_base_/datasets/coco_detection_dl.py and change:

data_root
ann_file
data_prefix

Duplicate configs/_base_/models/fast-rcnn_r50_fpn.py → configs/_base_/models/fast-rcnn_r50_fpn_dl.py and change num_classes to 1

Duplicate configs/faster_rcnn/faster-rcnn_r50_fpn_1x_coco.py and change to '../_base_/models/faster-rcnn_r50_fpn_dl.py','../_base_/datasets/coco_detection_dl.py'

Duplicate mmdet/datasets/coco.py → mmdet/datasets/deeplesion.py, rename the class to DeepLesionDataset. Configure the class to only include 'lesion', and leave only one arbitrary palette, removing the rest.

If this step is not done correctly, a ValueError: need at least one array to concatenate will occur

Follow the instructions at https://mmdetection.readthedocs.io/en/latest/get_started.html

If you have reached this point without changing the default path, it is convenient to directly move the current DeepLesion directory to mmdetection/data/ for use.

5️⃣ Run experiments

bash ./tools/dist_train.sh \
    {config_file_path} \
    16

Note

✉️ kwangrok21@naver.com

Ryoo72 / DeepLesion