ET | INF | MT | NES | IST | POL | BMT
Motivation
Coherent fiber bundle (CFB)-based endoscopes offer a minimally invasive imaging to access deep tissue in real time, such as brain tumor. Although honeycomb artifacts result in a low spatial resolution, deep neural networks (DNN) have shown high performance in reconstructing tissue features and enhancing image resolution. However, DNN-based computational imaging often raise concerns about the fidelity of the results.
In this work, an image reconstruction network framework with uncertainty estimation and confidence quantification is to be developed. Bayesian deep learning and Monte Carlo dropout approach will be used to build this framework for determining confidence map of the reconstructed fiber images. In this way, the reliability can be quantified and visualized with image reconstruction.
(a) Fiber endoscopy for brain cancer diagnostics
(b) Reconstruction with quantified confidence
Keywords: Deep learning, uncertainty estimation, image processing, Python, PyTorch
Tasks:
- Implementation of reconstruction DNN for CFB images
- Developing a method to model reconstruction uncertainty and quantify confidence
- Model optimization, result evaluation and documentation
Contact
Tijue Wang tijue.wang@tu-dresden.de