Toward a common validation methodology for segmentation and registration algorithms

Abstract

The National Library of Medicine and its partners are sponsoring Insight, a public software toolkit for segmentation and registration of high dimensional medical data. An essential element of this initiative is the development of a validation methodology, a common means of comparing the precision, accuracy, and efficiency of segmentation and registration methods. The goal is to make accessible the data, protocol standards, and support software necessary for a common platform for the whole medical image processing community. This paper outlines the issues and design principles for the test and training data and the supporting software that comprise the proposed Insight Validation Suite. We present the methods for establishing the functional design requirements. We also present a framework for the validation of segmentation and registration software and make some suggestions for validation trials. We conclude with some specific recommendations to improve the infrastructure for validating medical image processing research.