Enhancing Radiological Diagnosis: A Collaborative Approach Integrating AI and Human Expertise for Visual Miss Correction
![Screenshot 2024-01-20 at 1 53 56 AM](https://private-user-images.githubusercontent.com/30754423/298269593-67034d3b-70be-49f8-abd3-f1ea5ae9547c.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.A_8wrf3iRjvIJQYzfG7iEGTR0880_QQLCECUDcd7XGM)
https://drive.google.com/drive/folders/1h9ZoITAITS_mvGjyZi8dHUpo8dMn_tHz?usp=sharing
- Chexformer
- Spatio-Temporal Abnormal Region Extractor (STARE )
Orginally Chexformer is trained on the Chexpert Dataset and we provide the pretrained chexformer model below. We also provide a sample of dataset for training and testing the Chexformer on the below link
-
Chexformer model Pretrained model
https://drive.google.com/file/d/1SJeXGdqveZerVSfHEFRsxVo3TPY1FoId/view?usp=sharing
-
Chexformer Dataset [ Train & Val ]
https://stanfordaimi.azurewebsites.net/datasets/8cbd9ed4-2eb9-4565-affc-111cf4f7ebe2
python main.py --batch_size 16 --lr 0.00001 --optim 'adam' --layers 3 --dataroot data/
python main.py --batch_size 16 --lr 0.00001 --optim 'adam' --layers 3 --dataroot data/ --inference --saved_model_name=''
STARE module is trained on the combination of REFLACX and Egd-cxr. We provide the detailed discription and download link for the pre-processed dataset on the Data readme file. We provide the pre-trained model for STARE below
-
STARE ( Pretrained model ):
-
Sample Training Data (Not full )
-
Extracted Video spatial features: Frame features extracted by clipvit(spatial encoder ) can be downloaded below for the STARE module [ Train + Some Test samples]
https://drive.google.com/file/d/1rwNMLTfh0twaSlIqu9vY93OTFb1GT1kL/view?usp=drive_link
-
Ground Truth Summarized Reports: [Only Train]
https://drive.google.com/file/d/1io2FGE2IC1LtLmFWLO3yIB9BAq5SNKzL/view?usp=sharing
-
Input Summarized Reports:[Only Train]
https://drive.google.com/file/d/116pd5J2UZWykJqFZ4NLejxtZBkAD65ve/view?usp=sharing
-
python -m torch.distributed.launch --nproc_per_node 8 --use_env dvc.py --epochs=100 --lr=3e-4 --save_dir=vit --batch_size=2 --batch_size_val=2 --schedule="cosine_with_warmup"
To run the CoRaX on the Error dataset please use the Notebook uploaded above ## CoRaX-Errror_correction_final_file.ipynb ## . It uses the pretrained Chexformer and STARE which is provided on the below link. Please download from there.
https://drive.google.com/file/d/17fPfag9rBNmHMdFrwWIIfWUX6OKbjWZ9/view?usp=sharing
https://drive.google.com/file/d/1SJeXGdqveZerVSfHEFRsxVo3TPY1FoId/view?usp=sharing
Error Dataset with missing abnormalities mentioned in table-1. All the results are calculated on this error dataset
https://drive.google.com/drive/folders/1h9ZoITAITS_mvGjyZi8dHUpo8dMn_tHz?usp=sharing
https://drive.google.com/file/d/1XK9AoiXHKegUTQPK8mZW1VGfRCjySgnR/view?usp=sharing