Public defence in Computer Science, M.Sc. (Tech) Joel Jaskari

Title of the doctoral thesis: Machine Learning for Healthcare

Opponent: Professor Chris Holmes, University of Oxford / Alan Turing Institute, UK
Custos: Assistant Professor Arno Solin, Aalto University School of Science, Department of Computer Science

The public defence will be organised on campus.

The thesis is publicly displayed 10 days before the defence in the publication archive Aaltodoc of Aalto University.

Electronic thesis

Public defence announcement:

Classification of medical conditions and segmentation of tissues are currently performed manually by medical experts. Automation of these tasks could relieve the burden on medical experts. Data-driven machine learning approaches, especially those based on deep neural networks, have shown great promise in these tasks. However, the machine learning approaches need to be systematically analysed to estimate the reliability and robustness in practice.

This thesis studies machine learning for various tasks in healthcare using Finnish hospital datasets. Deep convolutional neural networks (CNNs) are used for classification of diabetic retinopathy from retinal images with clinically used severity scales. It is shown that even with a relatively small Finnish dataset of retinal images, the CNN can achieve comparable or better performance than previous works. In addition, Bayesian CNNs and a novel uncertainty measure are used to leverage uncertainty estimation in this task. The uncertainty estimates are shown to improve performance when used to refer uncertain images to experts. The CNN approach is also proposed for the segmentation of the mandibular canals from cone-beam CT volumes. It is shown to achieve state-of-the-art results and to perform the segmentation task with higher agreement to radiolgists than the between-radiologist agreement. Multiple machine learning approaches are studied for the early detection of neonatal mortality and morbidity. It is observed that the neonatal mortality and morbidities could be detected using as low as 12 hours of time-series data collected after the birth.

It can be concluded that Finnish hospital data can be used to train machine learning models for various tasks with high performance. These models could serve to augment medical professionals in their work by automation of tasks that require manual labour.