Rapid Interactive and Intuitive Segmentation of 3D Medical Images Using Radial Basis Function Interpolation

Abstract

Segmentation is one of the most important parts of medical image analysis. Manual segmentation is very cumbersome, time-consuming, and prone to inter-observer variability. Fully automatic segmentation approaches require a large amount of labeled training data and may fail in difficult or abnormal cases. In this work, we propose a new method for 2D segmentation of individual slices and 3D interpolation of the segmented slices. The Smart Brush functionality quickly segments the region of interest in a few 2D slices. Given these annotated slices, our adapted formulation of Hermite radial basis functions reconstructs the 3D surface. Effective interactions with less number of equations accelerate the performance and, therefore, a real-time and an intuitive, interactive segmentation of 3D objects can be supported effectively. The proposed method is evaluated on 12 clinical 3D magnetic resonance imaging data sets and are compared to gold standard annotations of the left ventricle from a clinical expert. The automatic evaluation of the 2D Smart Brush resulted in an average Dice coefficient of 0.88 ± 0.09 for the individual slices. For the 3D interpolation using Hermite radial basis functions, an average Dice coefficient of 0.94 ± 0.02 is achieved.

A survey on active learning and human-in-the-loop deep learning for medical image analysis

“…The methods discussed above have so far been concerned with producing segmentations for individual images or slices, however many segmentation tasks seek to extract the 3D shape/surface of a particular region of interest (ROI). Kurzendorfer et al (2017) propose a dual method for producing segmentations in 3D based on a Smart-brush 2D segmentation that the user guides towards a good 2D segmentation, and after a few slices are segmented this is transformed to a 3D surface shape using Hermite radial basis functions, achieving high accuracy. While this method does not use deep learning it is a strong example of the ways in which interactive segmentation can be used to generate high quality training data for use in deep learning applications -their approach is general and can produce segmentations for a large number of tasks.…”