Medical Image Detection

A detection framework for 3D medical images based on PyTorch refers to the modular design of MMDetection. The framework consists of one-stage and two-stage detectors. More methods, instructions, and experiment results will be updated.

Installation

Requirements

• g++ 7.5
• gcc 7.5
• cuda > 10.1
• torch > 1.6.0

MedVision

git clone https://github.com/TimothyZero/MedVision.git
cd MedVision
pip install .

Train

git clone https://github.com/JoeeYF/MedDetection.git
cd MedDetection
python train.py --config config_path
bash run.sh config_name 1 1 1

Supported Methods

Detectors

• Faster R-CNN
• Cascade R-CNN
• RetinaNet
• DeepLung
• CenterNet
• yolov4

Backbone

• ResNet
• ResNeXt
• SENet
• Res2Net

Neck

• FPN
• PAN
• BiFPN

DenseHead

• RetinaHead
• RPNHead

RoIHead

• BaseRoIHead
• DoubleHead
• CascadeHead

Other features

• OHEM
• 3D DCNv2
• 3D Soft-NMS

Dataset

Luna

The annotation json files are in COCO format.

Detection/Luna2016
├── train_images_test
    ├── subset0
        ├── ....nii.gz
    ├── subset1
    ├── subset2
    ├── ...
    ├── subset9
├── infer_dataset_0.json
├── infer_dataset_1.json
├── infer_dataset_2.json
├── ...
├── infer_dataset_9.json
├── train_dataset_0.json
├── train_dataset_1.json
├── train_dataset_2.json
├── ...
└── train_dataset_9.json

TODO

• more methods support
• experiment results
• dataset structure

References

MedVision

MMdetection

medicaldetectiontoolkit