Visual Analysis of Medical Image Segmentation Feature Space for Interactive Supervised Classification

Abstract

Classification of image regions is a crucial step in many image segmentation algorithms. Assigning a segment to a certain class can be based on various numerical characteristics such as size, intensity statistics, or shape, which build a multi-dimensional feature space describing the segments. It is commonly unclear and not intuitive, however, how much influence or weight should be assigned to the individual features to obtain a best classification. We propose an interactive supervised approach to the classification step based on a feature-space visualization. Our visualization method helps the user to better understand the structure of the feature space and to interactively optimize feature selection and assigned weights. When investigating labeled training data, the user generates optimal descriptors for each target class. The obtained set of descriptors can then be transferred to classify unlabeled data. We show the effectiveness of our approach by embedding our interactive supervised classification method into a medical image segmentation pipeline for two application scenarios: detecting vertebral bodies in sagittal CT image slices, where we improve the overall accuracy, and detecting the pharynx in head MRI data.