This is the source code of the paper “Automatic detection of oral cancer in smartphone-based images using deep learning for early diagnosis” (under review).
- The multi-class network is using the modified HRNet-W18, which achieves similar classification performance compared with the official code while reducing the computation complexity and the number of parameters.
- The code is a slightly modified version of the official code, by removing the BN&FC layers and changing the number of output channels in the representation head.
- For more details using HRNet, please go to the official website.
- Pytorch1.0
- CUDA 10.1
- Numpy
- Cv2
- Scikit-learn
- PyCm
python prepare_data/create_dataset.py- Note: This was only tested on Windows 10 system by us.
-
The training is only slight differences (with the same method but different super-parameters) on the original implementation.
-
The official pre-trained weights official website are also used. ( loading by the file of "save_weights/hrnetv2_w18_imagenet_pretained.pth" )
-
Prepare the dataset, the structure of files are as follows:
FOLDER/dataset/ ├── train │ ├── cancer │ ├── highrisk │ ├── lowrisk │ ├── normal │ └── ucler ├── valid │ ├── cancer │ ├── highrisk │ ├── lowrisk │ ├── normal │ └── ucler
python ./tools/train.py- The trained models (on oral dataset) are available at BaiduYun(Access Code:lyjv)
- The trained model should be under ./save_weights/
- The original capturing image (by smartphone camera) should be center-cropped (by create_dataset.py) before the test
python ./tools/test.py #predict and print results- The results were reported based on test results(test_result.csv), using the backend of PyCM and the Scikit-learn library
python ./tools/report_result.py