Plateforme d'évaluation multi-critères pour les algorithmes de traitement d'images médicales

Abstract

Each year, an increasing number of algorithms and methods are developed to segment anatomical structures given X-Rays. These algorithms provide assistance for clinicians diagnosis in order to interpret patients’ diseases. They, therefore, require to produce a segmentation as close as possible to reality. To ensure this, they are validated using performances tests applied on metrics, comparing the output to a reference. However, validation is very often subjective and each algorithm presents its own validation process. There is no standard protocol of validation and it is generally very challenging to compare an algorithm to others. Furthermore, in case of data taken on patients, validation are conditioned by the reference model, which is generally obtained by one clinician expert. In this thesis, we present a platform for evaluating imaging algorithms in medical images. This platform evaluation process has been divided in two major steps. In the first instance, we propose a way to obtain a valid reference : either by generating a gold standard on mathematical simulations ; either by using experts’ manual inputs and generating a bronze standard used for the reference. In the bronze standard case, we define a way of characterizing it using performances parameters. This allows obtaining a confidence interval of the reference and normalizing the subsequent algorithm analysis. Thereafter, we compare the behavior of algorithms to the reference using evaluation criteria defined by requirements. These criteria are an arrangement of metrics commonly found in the literature. Results are ultimately displayed in a radar style graph for performance analysis. These graphs offer a multi-criteria interpretation and are used to determine algorithms strengths and weaknesses. It is therefore possible to study many algorithms and compare them on a single graph that displays an easy to understand interpretation of algorithms. The developed platform is an interesting tool to not only pick the algorithm that best answers to specific requirements but also to have an idea on points that deserve to be improved on tested algorithms. This platform has been set up for 2D / 3D segmentation algorithms and for 3D reconstruction meshs. In the results section, we show an algorithm evaluation for some methods developed in the laboratory.

Date	24 Jan 2017
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jacques A. de Guise (Supervisor) & Nicola Hagemeister (Co-supervisor)