Automated confidence-based user guidance for increasing efficiency in interactive 3D image segmentation

In this thesis, we improve the standard 3D medical image interactive segmentation workflow. Drawing from the field of Active Learning, we propose a method for automating the process of deciding where the user should provide input next for optimally improving the segmentation. Specifically, we evaluate a given intermediate segmentation by constructing an uncertainty field over the image domain based on a multitude of segmentation quality metrics. We then find the plane that intersects with maximal uncertainty, and present it to the user for segmentation as an active batch query. We demonstrate the method through two embodiments, one using the Random Walker segmentation algorithm, and the other using the 3D Livewire method as seen in the software tool, TurtleSeg. We show that in both implementations, our method makes better decisions than user intuition and greatly reduces user interaction time.