Open junxiant opened 10 months ago
Hello there!
I haven't used the full dataset because of memory and time constraints. As most of the slices doesn't have any meaningful information about nodules, what I did was to use only the slices around the nodules. I'm guessing that a better approach would be to also include a few other images without nodules to make sure the module generalises better.
You can find more details about how I generated the dataset to use in the following notebook (it's a bit messy, sorry): https://github.com/nilsonsales/detr_luna/blob/master/detr/log_plots.ipynb
Can I check if the json annotations are in coco format too?
Another question which i would like to ask, how did you manage to deal with using grayscale images for training of this model? I saw that inside detr/datasets/luna.py, the normalization values are set to these
Hello, I am following your project and this looks interesting. Did you manage to train the DETR model on the LUNA16 dataset fully?