Master thesis presentation: Performance evaluation of feature learning for stroke classification in a microwave-based medical diagnostic system presented by Yimeng Hou

On Friday, September 18th, Yimeng Hou, MPCOM, will be presenting his master thesis Performance evaluation of feature learning for stroke classification in a microwave-based medical diagnostic system
Examiner: Tomas McKelvey

Abstract
In recent years, stroke has become an important issue to consider medically and socially since it poses high rate of human mortality and disability worldwide. It typically results from impaired blood supply in the brain. Pathologically, it can be classified into two types: ischemic stroke and hemorrhagic stroke. While early treatment can effectively save lives, the treatment to each type are different and wrong treatment may deteriorate patients’ condition. Therefore, it is necessary to diagnose the type of stroke before any treatment is performed. Medfield Diagnostics has developed a solution for stroke diagnostics in prehospital scenarios, a medical instrument called Strokefinder can be used to diagnose stroke type. This instrument uses microwave antennas to transmit and receive response signals and acquire the data for each patient. After that, machine learning algorithms are used to process the data and perform classification. Earlier methods already show high classification accuracy for predicting the correct type of stroke.

This thesis aims to investigate whether new methods would further improve the performance of stroke classification. Specifically, a methodology named feature learning will be evaluated. One of the common feature learning algorithms is autoencoder, which is a reconstruction-based unsupervised learning algorithm with the purpose of dimensionality reduction. Since the data from Strokefinder is massive and high-dimensional, applying autoencoder before a classifier will lower the dimension of acquired data and may possibly improve the classification performance.

The thesis starts with basic theory of autoencoder and support vector machine (SVM) classifier, introduces the complete pipeline for classification which includes preprocessing, autoencoder, SVM and performance evaluation. In addition to normal autoencoder, the idea of class-specific autoencoder will be brought up and implemented. Essentially, the performance of three schemes are compared in detail, which covers SVM without autoencoder, SVM with normal autoencoder and SVM with class-specific autoencoder. The performance results contain classification Accuracy and AUC with various cross-validation methods by testing stroke datasets from lab simulation in Medfield Diagnostics and other external datasets such as CIFAR-10 and MNIST. Results show the performance of combining feature learning with classifier could reach equivalent results comparing to case when using SVM solely. However, it shows the potential to classify massive dataset with high efficiency.

When?At 15.15 pm on Sep 18

Where?Blå Rummet (room 6414), Hörsalsvägen 11, 6th floor