Code for the Kaggle Pneumothorax Segmentation Challenge. Our team finished as 7th with a dice score of 0.8629. The following image shows the predictions of our final model. Non pneumothorax X-rays are shown in green. Predicted pneumothoraces are outlined in red.
The training data can be downloaded from here: https://www.kaggle.com/seesee/siim-train-test
The stage 1 test images and labels will be added to the training data.
Requirements:
pip install --upgrade pydicom tqdm opencv-python==3.4.5.20 albumentations==0.3.0 timm --userInstall apex for your system setup as explained here: https://github.com/NVIDIA/apex
$ ls | grep dicom
>>> dicom-images-test dicom-images-train$ cp Model_000_f00/*.py .
$ python train.py
$ cp Model_001_f00/*.py .
$ python train.py
$ cp Model_002_f00/*.py .
$ python train.pyThis will produce the weights Model_00*_f0*/*.pth and test predictions Model_00*_f0*/f0*-PREDS.zip.
$ python ensemble_all.pyYou can find a short overview of our entry on YouTube:
