Overview

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.

Setup

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 --user

Install apex for your system setup as explained here: https://github.com/NVIDIA/apex

$ ls | grep dicom
>>> dicom-images-test dicom-images-train

Training & Predictions

$ cp Model_000_f00/*.py .
$ python train.py
$ cp Model_001_f00/*.py .
$ python train.py
$ cp Model_002_f00/*.py .
$ python train.py

This will produce the weights Model_00*_f0*/*.pth and test predictions Model_00*_f0*/f0*-PREDS.zip.

Ensemble

$ python ensemble_all.py

Demo Video

You can find a short overview of our entry on YouTube: