Deep learning with weak annotation from diagnosis reports for detection of multiple head disorders: a prospective, multicentre study
This project contains source code and data for our manuscript "Deep learning with weak annotation from diagnosis reports for detection of multiple head disorders: a prospective, multicentre study".
- Python >= 3.6
- CUDA >= 10.0
- pip >= 20.0
We recommend that you create a new python environment by anaconda or virtualenv . Other requirements can be found in requirements.txt and installed by:
pip install -r requirements.txtTo visualize results on all four types of head disorders, run demo.sh :
source demo.shThe output images are saved in ./vis , and detailed output predictions are saved in ./stat_info . Here is an example from a CT scan (./data/0/0_0) with hemorrhage lesion:
The argument target_focus specifies the target head disorder:
| target focus | head disorder |
|---|---|
| 0 | hemorrhage |
| 1 | brain ischemia |
| 2 | skull fracture |
| 3 | tumor |
You can choose available GPU device by setting CUDA_VISIBLE_DEVICES in command line.
First, download dataset from http://headctstudy.qure.ai/dataset to somewhere ($PATH), then extract data by:
source make_cq500.sh $PATHafter this step, there should be extracted data in ./data/cq500 . Finally run test.sh :
source test.shthe AUC score should be printed on screen. The output of hemorrhage experiment looks like:
and the output of fracture experiment looks like:
Detailed predictions are saved in ./cq500_stat_info .
You can choose available GPU device by setting CUDA_VISIBLE_DEVICES in command line.
the computer-aided detection software can be installed directly with cad.exe
This project is only available to reviewers now.