Project: Federated Learning for detecting Lung Disease in Developing Countries
Description
Background
Delft Imaging develops mobile X-Ray machines that allow screening in low resource settings, such as developing countries or remote locations. Due to the lack of qualified professionals in these settings, they use computer vision models to automate screening of patients for diseases such as tuberculosis using images captured by these X-Ray machines. Specifically, their CAD4TB software is used daily in over 200 sites all over the world to screen patients for tuberculosis.
Due to privacy restrictions, Delft Imaging is interested in training these models using Federated Learning combined with privacy-preserving protocols such as differential privacy.
Project proposal
To develop a prototype full-stack FL system that runs on NUC boxes in developing countries, similar to for example this and this paper.
Developing such a prototype will require solving novel problems that can be turned into research questions that are still open and can be defined by the student, based on their interests.
Example focuses may be:
- privacy vs. performance trade-offs
- continual learning
- data selection / quality assessment
- federated learning on volatile and low-resource infrastructures systems
- model compression
Prerequisites
- Strong software engineering skills
- Familiarity with Computer Vision and Deep Learning
- Experience working with mini-computers such as NUCs, Raspberry Pis is helpful.
- Familiarity with Federated Learning and privacy-preserving machine learning is helpful.
Practical details
- Duration of at least 4 months
- At least 3 days at the office in 's-Hertoghenbosch
- € 625 gross reimbursement per month
Details
- Supervisor
-
Tim d'Hondt
- Secondary supervisor
-
Mykola Pechenizkiy
- External location
- Delft Imaging Systems ('s-Hertogenbosch)
- Interested?
- Get in contact