Research Information System
Multimedia Tools and Applications, 2025 (SCI-Expanded)
Accurately delineating both livers and liver tumors holds crucial significance in the diagnosis of liver cancer. Deep learning approaches have demonstrated superior performance compared to conventional methods in achieving precise segmentation outcomes. However, they often fall short in obtaining satisfactory results because of several factors, such as very small tumors, blurred image boundaries, and noise in the images. Therefore, in this work, an efficient U-Net-shaped and hybrid attention-aware network has been developed for segmentation of both livers and liver tumors. Its effectiveness has been obtained by: (i) dense and improved residual connections utilized to strengthen feature fusion and propagations in the architecture, (ii) a transformer unit constructed with a fusion of global and local feature attention mechanisms to capture both global (long-range) and local dependencies, (iii) the usage of leaky rectified linear units and atrous separable convolutions to provide efficiency in feature extraction. To make fair comparative evaluations, recent methods proposed for liver or/and liver tumor segmentations have been implemented using the same datasets, and their abilities have been assessed based on the same measurements. The main contributions of this paper are fivefold: (i) Introducing a novel network model effectively designed by integrating enhanced dense connections, transformer blocks, patch combining, and expansion layers. (ii) Introducing an improved mobile-inverted bottleneck convolution block for feature extraction. (iii) Presenting the effectiveness of the proposed transformer unit and enhanced feature linkage modules separately. (iv) Demonstrating the architecture's performance in the segmentation of both livers and liver tumors, indicating its superiority with a minimum improvement of 1.53% and 10.4%, respectively, in terms of Dice coefficients. (v) Providing fair comparisons of recent methods using the same metrics after implementation with the same datasets.