This repo contains my solution to the SIIM-ACR Pneumothorax Segmentation; Identify Pneumothorax Disease in Chest X-Rays challenge. The challenge consists on detecting pneumothorax disease (i.e. collapsed lung disease) from x-ray images and masking the location of the disease within the x-ray image.

1. The first step is to run the images through an ensemble of EfficientNet (Tan & Le 2020) based binary image classifiers to determine whether the x-ray presents pneumothorax disease.

2. The second step, runs the image through an ensemble of Unet (Ronneberger et al., 2015) and Unet ++ (Zhou et al., 2019) networks with EfficientNet backbones to identify the location of the disease.

This approach achieves a Dice coefficient of 0.8522 in the offcial private test data of the competition, on par with the top 3% of the results.

Re-running the Notebook

• Open the notebook in Colab and select TPU as accelerator (recommended).
• Update the GCS Path as indicated in the notebook
• Run all

To train using TPUs, I extracted the original data from DICOM files (X-Rays, patient metadata) and CSV files (mask RLE encodings) and placed it in TFRec files. The original DICOM and CSV data can be found here. The TFRecs used in the notebook can be found here.

Thanks to The Society for Imaging Informatics in Medicine (SIIM) and the American College of Radiology (ACR) for creating and providing the dataset.

• Ronneberger O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv:1505.04597v1.
• Tan, M., & Le, Q. V. (2020). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv:1905.11946v5.
• Zhou, Z., Siddiquee, M., Tajbakhsh, N., & Liang, J. (2019). UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation. arXiv:1912.05074v2.