Multi-Criteria Evaluation Framework for Deep Learning Architectures in Medical Image Segmentation

Abstract

The rapid expansion of deep learning architectures for medical image segmentation presents a major challenge to selecting optimal model to implement in practice. While several traditional evaluation methods have been proposed to solving this selection problem, they cannot use criteria interdependencies and expert judgment   to select deep learning architectures for medical image segmentation. This study, therefore, presents a multi-stage multi-criteria decision-making framework to address this knowledge gap. The framework contains an improved DEMATEL (Decision Making Trial and Evaluation Laboratory), intuitionistic fuzzy AHP (Analytic Hierarchy Process), and adaptive VIKOR (VIekriterijumsko Optimizacija Kompromisno Resenje). This study evaluated framework performance using six deep learning architectures and thirteen criteria. The DEMATEL-fuzzy AHP results showed sensitivity, specificity, and accuracy the most important criteria for evaluate the deep learning architectures for medical image segmentation. The enhanced VIKOR model identified Swin-Unet and nnU-Net as the optimal compromise and best solution, respectively, for the optimum compromise. Based on sensitivity analysis that was conducted, the VIKOR ranked architecture as UNETR >> nnU-Net >> Attention U-Net >> TransUNet >> Swin-Unet >> classic U-Net. This study’s findings have showed that the proposed framework can be used to support deep learning architectures for medical image segmentation decisions.