Deep Learning for stroke detection.
When a stroke occurs, time is of the essence! The time it takes for the patient to the hospital is the absolute main factor in the outcome of the stroke. About half of the people that suffer from stroke gets facial paralysis (loss of muscle control in one side of the face) as a symptom. ResQ Biometrics is collecting images and videos of stroke patients and developing an API to predict and warn people when they are experiencing a stroke. To support this goal we are looking for master students for two different projects.
Personal Prediction Model – automatic fine-tune deep learning model: Human faces come in a lot of varieties and developing machine learning models that must generalize to every user is a challenging task. Utilizing person-specific training/configuration of the models should (will?) increase the accuracy by a large margin. We want to investigate how to transfer the learning of a general model to a person-specific one. The goal of the project would be a method/approach on how to automatically fine-tune deep learning models to a specific user. The approach should deal with storage space limitations for multiple users.
Deep anomaly detection on image/video data for stroke detection: Deep learning algorithms need a lot of images or videos to predict with high accuracy. In medical applications, the data is often scarce and/or is expensive to collect. An alternative to learning from the data distribution one is interested in is learning from ‘the other data’. Deep Anomaly Detection (AN) used in this context, is an approach that requires significantly less data domain-specific to train than more classical techniques like CNN’s. We would like to investigate/research the use of AN to predict stroke from images/videos. The result could be an approach that works for our domain.
- Michael Riegler
- Pål Halvorsen
- Gisle Halvorsen, ResQ Biometrics
- Programming knowledge. Preferably in Python or another scientific programming language like R or MATLAB
- Basic machine learning knowledge
- Some hands-on experience with a machine learning framework (like Tensorflow/Keras, PyTorch, ++) is a plus.