Comparative Analysis of Existing UNet-Based Medical Image Segmentation Models and an Optimized ResAttTransUNet Framework

Abstract

Despite the popularity of medical image segmentation in computer-aided diagnosis (CAD), the choice of an effective architecture for deep learning is still a difficult task as the datasets, network architectures, activation functions and the assessment approaches are different. In this work we compare in detail the segmentation models based on UNet recently developed and we propose a new architecture based on UNet, called ResAttTransUNet, for improving the segmentation accuracy, robustness and generalization across several imaging modalities. In the proposed model, the UNet architecture is used to integrate the residual blocks, dual attention mechanisms and transformer bridge, while the effects of GELU, PReLU and Leaky ReLU activation functions are evaluated. A set of experiments were performed on three benchmark datasets, namely the brain MRI dataset, BRATS2020, the liver CT data, LiTS17, and the breast ultrasound data, BUSI, with a common experimental setup. The Dice coefficient, Intersection over Union (IoU), qualitative analysis, ablation studies, and comparative evaluation were used to evaluate the performance. The results show that GELU always performed best in all datasets. The proposed ResAttTransUNet achieved the best performance on LiTS17 with Dice score and IoU of 0.9619 and 0.9299, respectively, and 0.8278/0.7470 on BRATS2020 and 0.7466/0.6074 on BUSI. The proposed method shows an excellent performance in segmenting multimodal medical images and also highlights its segmentation accuracy and generalization capabilities over the existing state-of-the-art methods, by making optimal use of the residual learning, attention mechanism, transformer-based modeling, and optimization of the activation functions.