This repository contains the code to solve the Multimodal Brain Tumor Segmentation Challenge 2017 as part of the seminar "Practical Video Analysis" during the summer term 2017 at Hasso Plattner Institute for Software Systems Engineering. Most of the model is based on pix2pix-tensorflow.
- The introduction tasks can be found inside the brain_health_classification folder. We familiarized with the BraTS challenge and Deep Learning by first using simple Keras models to distinguish between healthy and unhealthy brains.
- BraTS related tasks
- The classification algorithm can be found inside the classification folder. The task given by the supervisor demanded to distinguish between low- and high-grade gliomas. We used a vanilla CIFAR10 model which was able to distinguish between the mentioned tumor types with ease.
- The segmentation algorithm can be found inside the pix2pix-ultimate folder. The approach is based on pix2pix-tensorflow and uses a Conditional Generative Adversarial Network to distinguish between real segmented brain images and fakes ones.
- The survival rate prediction can be found inside the survival_rate folder. The preferred approach uses a CNN to predict the survival rate of the patient. The CNN model outperforms the Support Vector Regression model in precision as well as variance of the error.
The demo code is located in the server directory. It requires a trained segmentation model.
