Automated bone erosion scoring for rheumatoid arthritis with deep convolutional neural networks
In fall 2017 I wrote a project thesis at the Zurich University of Applied Sciences, where I examined whether bone erosion scores of patients with rheumatoid arthritis can be predicted wih deep convolutional neural networks. The networks were trained on cropped x-ray images of left hands. The code in this repository was used to obtain the results in the thesis.
The thesis can be found here: /doc/project.pdf
All jupyter notebooks can be run on the following docker container: tensorflow:1.4.0-gpu-py3
Below is a list of the files in the master branch of this repository with a description of what they are used for. There is also the model_selection branch which contains the other models which were not selected.
| Filepath | Description |
|---|---|
| /correlation_analysis/correlation_analysis.ipynb | Jupyter Notebook that shows correlations between the Rau-score and the DAS-score |
| /correlation_analysis/plots_for_thesis.Rmd | R-Markdown file used to create the correlation plots for the thesis |
| /doc/img/ | This folder contains all images used in the thesis |
| /doc/project.pdf | The thesis |
| /doc/project.tex | The LaTex file for the thesis |
| /doc/project.tex | The BibTex file with the references of the thesis |
| /tensorboard/ | This folder contains all the tensorboard logs from the training of the models |
| /tsne/tsne_regression.R | This R-script contains an analysis of the outliers in the T-SNE |
| attention_map_classification.ipynb | Jupyter notebook that shows the attention map of the classification model |
| attention_map_regression.ipynb | Jupyter notebook that shows the attention map of the regression model |
| deepxray_classification_weights.ipynb | Jupyter notebook used for the training of the classification model with weighted loss function |
| deepxray_classification_weights_transfer_learning.ipynb | Jupyter notebook used for the training of the transfer learning classification model with weighted loss function |
| deepxray_regression_original.ipynb | Jupyter notebook used for the training of the regression model on original data |
| deepxray_regression_original_transfer_learning.ipynb | Jupyter notebook used for the training of the transfer learning regression model on original data |
| embeddings_classification.ipynb | Jupyter notebook with T-SNE of the embeddings of the classification model |
| embeddings_classification_transfer_learning.ipynb | Jupyter notebook with T-SNE of the embeddings of the transfer learning classification model |
| embeddings_regression.ipynb | Jupyter notebook with T-SNE of the embeddings of the regression model |
| embeddings_regression_transfer_learning.ipynb | Jupyter notebook with T-SNE of the embeddings of the transfer learning regression model |
| prediction_time.ipynb | Jupyter notebook that loads the two models and creates predictions. Measures the execution time for both predictions. |
| preprocessing.ipynb | Jupyter notebook that preprocesses the data (train, test & validation set of images and labels) for the classification model |
| preprocessing_regression.ipynb | Jupyter notebook that preprocesses the data (train, test & validation set of images and labels) for the regression model |
| validate_classification.ipynb | Jupyter notebook with predictions of the classification model for the test set |
| validate_regression.ipynb | Jupyter notebook with predictions of the regression model for the test set |