For this Lung Cancer detection project, we used the radio library. The radio library can be downloaded from here Read radio's documentation at here
Also, read the (requirements) file on the needed python packages to run this project.
We used the Luna16 Data set to train and test our model. This dataset consists of 10 submodules and an annotation.csv. The datasets are 5-7GB in size each, For our implementation, you can use any number of subsets you wish. The trade-off is, less subsets = faster runtime and worse results. More subsets = slower runtime but better results.
Begin by setting up a file structure in the project folder for the data crop dumps. Out prefered file setup was a data directory for saving the subsets. In that data directory, we created subfolders for the cancerous and noncancerous data. The directories structure was: annotations --> data/ subsets --> data/ cancerous --> data/luna_split/cancerous/ noncancerous --> data/luna_split/noncancerous/
Replace these paths with the path above
nodules = pd.read_csv('YOUR_ANNOTATIONS_PATH')
lunaix = ds.FilesIndex(path='SUBSETS_PATH', no_ext=True)
cancer_path='CANCEROUS_PATH', non_cancer_path='NON_CANCEROUS_PATH'
DIR_CANCER = 'CANCEROUS_PATH/*' DIR_NCANCER = 'NONCANCEROUS_PATH*'
You then can run the module and wait for the outputs.
Hyperparamters can be adjusted here:
train_time = 1 my_epoch = 1 bs = 10 loss_his = []
If the notebook isnt displaying at the moment and says: "Sorry something went wrong. Reload?" view the notebook at this link