You can see our final report and presentation for details.
-
Dataset: IU X-ray, MIMIC-CXR (Actually, we used pretrained CLIP model with the dataset)
-
Model architecture for training: ResNet18, EfficientNet-b4
-
Loss functions for training: Cross Entropy Loss, CLIP directional Loss
-
AdamW optimizer and 2e-3 learning rate
-
Result:
-
Reference:
https://github.com/atimashov/cxr-report-generation (our baseline),
https://github.com/rajpurkarlab/CXR-RePaiR (CLIP model),
https://github.com/rinongal/StyleGAN-nada (CLIP directional Loss)