Individual projects can be selected on the project’s portal.
- Web-based user-interface for the exploration of brain data: The aim of this project is to build a web-based 3D exploration and labelling user interface to make this highly complex data exportable for the general public. Advanced view management techniques (e.g. https://www.youtube.com/watch?v=XUQu4XAc2MI), disassembly techniques (e.g. explosion diagrams, https://www.youtube.com/watch?v=hXoeMdTx48s), and interaction techniques (e.g. using Intel’s RealSense https://www.youtube.com/watch?v=uINRC83tlTA), and real-time data visualisation techniques are to be explored in the course of this project.
- Deep-learning assisted semi-automatic, interactive 3D segmentation of medical data: A goal of this project is to develop a generally applicable method for 3D segmentation of medical images that is able to train extremely fast on new data and to provide tissue segmentations in real-time. Initially this approach will be tested on thoracic CT scans for the semi-automatic, assisted segmentation of lungs and lung nodules.
- GPU Accelerated SCNNs for the Prediction of Breast Cancer Proliferation: The aim of this project is to evaluate sparse CNNs for the prediction of breast cancer proliferation in histopathological images.
- Massively parallel slice-to-volume registration for motion compensation in MRI: The aim of this project is to implement rigid slice-to-volume registration for slices of arbitrary size using massively parallel computing platforms like modern GPUs. The developed method will extend an existing platform for SVR (https://github.com/bkainz/fetalReconstruction), based on Nvidia’s CUDA GPU programming language. Test data and appropriate, remotely accessible hardware will be provided in the computing lab.
- Applications for pocket-sized supercomputers: The aim of this project is to test and review the capabilities of pocket-sized supercomputers like the parallella platform (http://www.parallella.org/) or modern mobile phones and mobile GPU platforms for real-time applications with low engery demands. Computers have come a long way in 40 years, and mobile computers have come the farthest in the shortest time – some of them are more powerful than early supercomputers. A tweaked Motorola Droid can hit speeds more than 15-times faster than the Cray-1 supercomputer of 1979. Apple’s iPad 2 rivals the Cray 2 supercomputer, the world’s fastest computer in 1985. What are promising applications of such platforms for the future?
- Deep bottlenecks: The goal of this project is to develop a framework that allows to evaluate the computational bottlenecks of deep learning approaches and to propose improvements automatically. This method should be generally applicable. Initial exemplars may use natural images and medical image data.
- Deep introspection: The goal of this project is to develop a framework that allows to intuitively explore trained deep network models based on state-of-the-art deep learning methods. This method should be generally applicable. First exemplars may use natural images and medical image data.