Pneumothorax-Segmentation

Top 7% Place Solution for SIM-ACR Pneumothorax Segmentation competition on Kaggle

My model was an enhanced version of Unet with EfficientNet B4 encoder and Resnet decoder on 512x512 images. The model scored 0.8413 LB (dice loss) on Private Leaderboard.

You can also find the code of an ensembling model that has a higher score 0.8435 LB on Private Leaderboard.

Flowchart

Single Model Architecture

Requirements

Download the weights

Download the corresponding weights and place them in the current folder on the Git repository.

For the single model: weights

Set up the Environment

• You will need Python3 with the following librairies and Tensorflow v1 (v1.14.x or v.1.15.x works)

• run: pip3 install -r requirements.txt

Work report

You can have access to my Work Report of this competition in the PDF file Work-report.pdf. You can find here many information about my model.

Instructions

• Donwload this preprocessed dataset of the competition: dataset
• Download the weights and set up your environment.
• Run the notebook pneumothorax_segmentation_model.ipynb

You may reduce the batch size in the hyperparameters of the model if you don't have enough GPU memory.