Public defence in Computer Science, M.Sc. (Tech) Jaakko Sahlsten

Title of the doctoral thesis: Applicability and Robustness of Deep Learning in Healthcare

Doctoral student: Jaakko Sahlsten
Opponent: Prof. Frank Emmert-Streib, Tampere University, Finland
Custos: Prof. Jouko Lampinen, Aalto University School of Science, Department of Computer Science

Healthcare professionals analyse medical images in a time-consuming, laborious, and routine clinical tasks that can be expected to increase with population ageing. Recently, deep learning methods, especially convolutional neural networks (CNNs) have been performing at the state-of-the-art level in various image analysis tasks. However, these methods can be overconfident and their performance may reduce when evaluated on out-of-distribution datasets e.g., patients imaged with other devices. 

In this thesis CNNs were developed for the clinical image-based tasks of diabetic retinopathy classification, mandibular canal localisation, organs at risk segmentation, and oropharyngeal cancer segmentation. The methods were evaluated for their applicability and robustness for the task by analysing the performance of the method with multiple diverse datasets. In addition, Bayesian CNNs and novel uncertainty measures were developed and evaluated for uncertainty utilisation in clinical workflow e.g., by referring to most uncertain cases to healthcare professionals while automatically analysing the rest. 

The applied CNN methods turned out to have clinically acceptable or medical expert level performance in the diabetic retinopathy classification, mandibular canal localisation, and oropharyngeal cancer segmentation using the in-distribution datasets. Moreover, the change in performance of the methods varied when evaluated on the out-of-distribution datasets depending on the task and dataset. Furthermore, the Bayesian CNNs and novel uncertainty measures, both improved the utility of uncertainty but the specific recommendations for the choice of method remain unclear. 

The findings in this thesis motivate expanding the analysis of CNNs in other medical imaging-based tasks based on the high level performances. However, to improve the trustworthiness of the methods, additional analysis using diverse datasets and improved utility of uncertainty through the use of Bayesian CNNs and task specific uncertainty measures may be warranted.

Thesis available for public display 10 days prior to the defence at: https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/

Doctoral theses in the School of Science: https://aaltodoc.aalto.fi/handle/123456789/52