Master's Thesis

The big finale - my master thesis. Over a year I had the opportunity to work at the Institute of Visual Computing & Human-Centered Technology at the University of Technology Vienna with incredible people. My project and later thesis has the title "Multi-faceted Visual Analysis of Inter-Observer Variability". From this project we wrote a paper that was submitted to IEEE Computer Graphics and Application, which in case it is accepted will be published in a special issue coming out in September.

Abstract

Despite the advancements in auto-segmentation tools, manual delineation is still necessary in the medical field. For example, tumor segmentation is a crucial step in cancer radiotherapy and is still widely performed by hand by experienced radiologists. However, the opinions of experienced radiologists might differ, for a multitude of reasons. In this work, we visualize the variability originating from multiple experts delineating medical scans of the same patient, known as inter-observer variability.
The novelty of this work consists of capturing the process of segmenting a target object. The focus lies in gaining insight into the observer's thought processes and reasoning strategies. To investigate these aspects of segmenting we conduct a data acquisition with novice users and experts, capturing their thoughts in a think-aloud protocol and their areas of attention by tracking their mouse-movement during the segmentation process. This data is visualized with our Multi Observer Looking Environment (MOLE). MOLE allows to gain deep insight into the observers' segmentation process and enables to compare different segmentation outcomes and how these occurred. With our proposed visualization techniques we emphasize regions of uncertainty that need more attention when delineating. Additionally, relevant keywords are extracted from the think-aloud protocol and aligned with the positions in the segmentation, providing information about the thought process of an observer. We link the initial image to a three-dimensional representation of the delineations and provide more details of the think-aloud protocol on demand.
Our approach is universal to segmentation, attention and thought process data regardless of the domain of the data. We show how MOLE can be used with a medical dataset as well as an artificially created dataset. By validating our approach with the help of a medical expert actively working in the field, we define potential use cases in the existing pipeline of tumor delineation for cancer treatment.

Have a look at the source code right here. You can find the thesis here.