Amose²una metodología de caracterización del error en segmentación de objetos amorfos. La segmentación de hiperintensidades cerebrales como caso de estudio

Abstract

The design process of an object segmentation system usually follows an iterative and incremental development pattern where experts analyse, in the evaluation stage, the errors made by the system, in order to know them and propose solutions for the refinement of the system. Although there is a set of well-defined metrics to evaluate the quality of the segmentation in an automatic way, the evaluation of the error with the objective of finding its cause has always been a very handcrafted task. Therefore, new methods and tools are needed to facilitate the evaluation of a system during its design, which must be explainable, flexible, interactive, analytical and contextual. In this work, with the aim of discovering new knowledge of the error in an automated way, a methodology is proposed for the characterization of the error in segmentation of amorphous objects under the hypothesis that there are a reduced number of patterns that explain a good part of the errors. This methodology, named AMOSE2, proposes to model the error by individual objects and performs a detailed description of them by means of a vector of characteristics for each error object, which allows a deeper analysis of the segmentation errors using artificial intelligence clustering and outlier detection techniques. In addition, knowledge models are introduced using ontologies to describe the visuo-spatial (reusable across domains) and contextual (domain specific) features of the segmented objects and their errors. Two types of features are distinguished according to the origin of the information, those generated from external information to the segmentation system and those generated internally by the segmentation system (only accessible in grey box systems). Their ontological description allows to improve the manipulation of variables, to facilitate the interaction with the expert or to make a feature selection based on both statistical and semantic measures. The detected error patterns are analysed by the system in order to select the most relevant ones, which will allow to guide the efforts in the next refinement cycle in the design stage. A prototype has been implemented following the definitions of the methodology, consisting of an error object segmentation description and analysis tool (AMOSE2 analysis) and a tool for interactive exploration of results (AMOSE2 web report). For its validation and example of use, a real problem has been used, the segmentation of cerebral white matter hyperintensities in magnetic resonance images. Three systems have been evaluated on five different datasets to show the two ways of using the AMOSE2 methodology: the individual analysis, where a proposed segmentation system is compared with respect to a given reference segmentation, and the comparative analysis, where it is as well compared with respect to another segmentation system. The findings of relevant error patterns have been useful for refining the AMOS-2D segmentation system.