Conventional transformation methods primarily rely on the intra-domain knowledge. On the other hand, advanced generative adversarial networks generate inter-domain images lacking variety, particularly in small-scale datasets. In light of these challenges, we propose that accurately controlling the variation degrees of synthetic samples can reshape the decision boundary in hyperplane space for the downstream classifications.
We developed a DisGAN effectively controlling the variation degrees of generated iner-domain and intra-domain samples. The locations are respectively defined using the vertical distances of the inter-domain samples to the optimal hyperplane and the horizontal distances of the intra-domain samples to the source samples, which are determined by Hinge Loss and Pythagorean Theorem.
#--------------------Breast Ultrasound--------------------# #-----------------------COVID-CT-----------------------#
- Linux or Windows
- Python 3.8+
- NVIDIA GPU + CUDA CuDNN
- anaconda virtual environment
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia conda install tqdm conda install matplotlib==3.3.4 conda install seaborn conda install scikit-learn
covid_ct：
├─train├─X
│ │ 001non-covid19.png
│ │ 002non-covid19.png
│ │ 003non-covid19.png
│ ├─Y
│ │ 001covid19.png
│ │ 002covid19.png
│ │ 003covid19.png
├─valid├─X
│ │ 001non-covid19.png
│ │ 002non-covid19.png
│ │ 003non-covid19.png
│ ├─Y
│ │ 001covid19.png
│ │ 002covid19.png
│ │ 003covid19.png
├─test ├─X
│ │ 001non-covid19.png
│ │ 002non-covid19.png
│ │ 003non-covid19.png
│ ├─Y
│ │ 001covid19.png
│ │ 002covid19.png
│ │ 003covid19.png
python train.py --cuda --dataset_name covid_ct --project_name cyclegan_convnext_tiny --aux_classfier pretrained_aux_convnext_tiny.pth
--model_name convnext_tiny --lambda_vertical 0.01 --lambda_horizontal 0.01
python test.py --cuda --dataset_name covid_ct --project_name cyclegan_convnext_tiny --aux_classfier pretrained_aux_convnext_tiny.pth
--model_name convnext_tiny --classifier downstream_convnext_tiny.pth