Predicting Position and Volume of Hemorrhagic Strokes
Examensarbete för masterexamen
Complex adaptive systems (MPCAS), MSc
This study has explored different neural network methods for position and volume prediction of hemorrhagic strokes. Three different data sets of microwave data were used as input for the different networks. A simple diagnostic classifier was used as benchmark to help in evaluating success. The two largest challenges of the study was instrument variations in the data as well as the limited data available. A Multi- Source Adversarial Domain Adaptation (MSADA) network was introduced to lower the effect of the instrument variations, and a Divergence Based Domain Adaptation (DBDA) network was implemented to attempt to resolve the limited number of data samples. The networks showed promising results for both position and volume in all three data sets used. The MSADA network successfully lowered the instrument specific noise when predicting volumes, but was concluded to be unnecessary for position classification. The DBDA network was not enough to remedy the lack of sufficient data.
Deep neural networks , machine learning , adversarial domain adaptation , multi-source adaptation , microwave data , stroke detection