Prediction for 2018 Data Science Bowl challange.
Semantic segmentation model with UNet architecture using Keras. To get data and more ifnormation can be found at Kaggle 2018 Data Science Bowl challange
Once downloaded unzip relavent archives so that structure looks like this: data-science-bowl-2018 |__Basic_EDA.ipynb |__predict_masks.py |__stage1_sample_submission.csv |__stage1_solution.csv |__stage1_test __|__0114f484a16c152baa2d82fdd43740880a762c93f436c8988ac461c5c9dbe7d5 _____|__images ________|__0114f484a16c152baa2d82fdd43740880a762c93f436c8988ac461c5c9dbe7d5.png ... |__stage1_train __|__0a7d30b252359a10fd298b638b90cb9ada3acced4e0c0e5a3692013f432ee4e9 ____|__images ______|__0a7d30b252359a10fd298b638b90cb9ada3acced4e0c0e5a3692013f432ee4e9.png ____|__masks ______|__0adbf56cd182f784ca681396edc8b847b888b34762d48168c7812c79d145aa07.png ...
For preprocessing images are resized to the same squares (128x128 pixels by default). For further data augmentation one might apply other preprocessing and transformations (blurring, mirroring, etc).
The model replecates the basic structure of U-Net model with reduced number neurons which still yields adequate predictions.
Scripts where developed and ran using Google Colab Notebook (Basic_EDA with Jupyter Notebook). It was not tested localy (due to technical difficalties with tensorflow backend). Replaced behavior can be obtained by running train script from Google Colab.