PhD defence R. (Richard) Bortsov

Dissertation in short:

In this thesis, I attempted to address two problems.

The first problem was improving accuracy of deep learning methods for medical image analysis. This problem is the subject of Part I of the thesis, which contains chapters 2, 3, and 4. In Chapter 2, I developed an algorithm to automatically detect and measure the volume of intracranial calcifications in non-contrast head computed tomography (CT), which is a promising imaging marker for the assessing risk of cerebrovascular diseases including clinical stroke. In Chapter 3, I developed a novel algorithm for automatic estimation of severity of emphysema, one of the lung diseases that comprise chronic obstructive pulmonary disease (COPD), in chest CT. In Chapter 4, I proposed a novel algorithm for learning from data containing images with labels and without labels, i.e. an algorithm using partially labeled data.

The second problem, which is the focus of Part II of the thesis, including chapters 5 and 6, was assessing vulnerability of deep learning methods for medical image analysis to adversarial examples—inputs intentionally designed to manipulate outputs of these methods. In Chapter 5, we studied factors that affect effectiveness of adversarial attacks on medical image analysis systems. In Chapter 6, I studied so-called targeted adversarial attacks: the attacks that do force the target algorithm to produce a specific, rather than arbitrary, incorrect output.