Closed kimwj94 closed 6 months ago
Hi, may I ask how did you get the class label of the 4000+ augmented bone suppression data? Thank you.
I am not sure why you need the class labels for bone suppression. We train the models on bone-suppressed and non-bone-suppressed image pairs so the model predicts a bone-suppressed image when a non-bone-suppressed image is fed into it.
Hi, thank you for your research.
Bone Suppression Dataset: The researchers from the Budapest University of Technology and Economics used their in-house clavicle and rib–shadow removal algorithms to suppress the bones in the 247 JSRT CXRs and made the bone-suppressed soft-tissue images publicly available at https://www.mit.bme.hu/eng/events/2013/04/18/boneshadow-eliminated-images-jsrt-database. The link is currently broken. We have augmented this dataset to create 4000+ original-bone-suppressed image pairs to train our proposed bone suppression models. The data is available at https://drive.google.com/drive/folders/1m4hlwglZIK14Mlkjf3YsNHRfXLJlfbBN?usp=sharing. Please cite our study if using these data and codes for your research:
I can get 4000+ augmented bone suppression data from the google drive link, but can I get original 247 CXR dataset that before augmented?
I need original dataset because when I tried to split 4000+ augmented dataset into train/val/test datasets, differently augmented images from same original image go to both train set and test set. That causes abnormally high test accuracy.
Thanks for your interest in our study. The JSRT dataset (247 CXR images) can be downloaded from http://db.jsrt.or.jp/eng.php. There is also a Kaggle source https://www.kaggle.com/datasets/raddar/nodules-in-chest-xrays-jsrt?select=images where the data is available. Hope this helps.
Hi, thank you for your research.
I can get 4000+ augmented bone suppression data from the google drive link, but can I get original 247 CXR dataset that before augmented?
I need original dataset because when I tried to split 4000+ augmented dataset into train/val/test datasets, differently augmented images from same original image go to both train set and test set. That causes abnormally high test accuracy.