Generalisation of Segmentation Using Generative Adversarial Networks

IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1–4

Abstract

State-of-the-art deep learning algorithms are easily biased and evaluated in misleading scenarios, especially in the medical context, where scenarios change rapidly and diseases develop quickly. The BraTS 2024 GoAT challenge aims to evaluate how brain tumour segmentation algorithms can adapt to different circumstances when these are not available for training. Our solution utilises state-of-the-art conditional generative adversarial networks to generate realistic new cases and train a segmentation algorithm that takes advantage of the convolutions and attention mechanisms. Our solution achieved a DSC value of 0.855, 0.863, 0.883 and an HD95 value of 24.83, 24.10 and 21.72 for the enhancing tumour, the tumour core and the whole tumour in the validation set, respectively.

Authors 6

1. André Filipe Ferreira

   Essen University Hospital

   Affiliation as printed

   University Hospital Essen (AöR),Institute for AI in Medicine (IKIM),Essen,Germany

2. Gijs Luijten

   Essen University Hospital

   Affiliation as printed

   University Hospital Essen (AöR),Institute for AI in Medicine (IKIM),Essen,Germany

3. Behrus Puladi Aachen Institute of Medical Informatics

   Universitätsklinikum Aachen · RWTH Aachen University

   Affiliation as printed

   University Hospital RWTH Aachen,Institute of Medical Informatics,Aachen,Germany

4. Jens Kleesiek

   Essen University Hospital

   Affiliation as printed

   University Hospital Essen (AöR),Institute for AI in Medicine (IKIM),Essen,Germany

5. Victor Alves

   University of Minho

   Affiliation as printed

   University of Minho,Center Algoritmi / LASI,Braga,Portugal

6. Jan Egger

   Essen University Hospital

   Affiliation as printed

   University Hospital Essen (AöR),Institute for AI in Medicine (IKIM),Essen,Germany