E-TransConvNet: An enhanced transformer and convolutional network for medical image segmentation from ultrasound and CT images

Medical image segmentation in ultrasound and CT images plays a crucial role in the prevention, diagnosis, and treatment of various diseases and cancers. Convolutional neural networks (CNNs) are computationally efficient for medical image segmentation but struggle to capture long-range dependencies, which are essential for global contextual understanding. Transformers address this limitation through self-attention mechanisms. However, architectures composed solely of transformer blocks impose prohibitive computational costs and compromise local spatial continuity. To overcome these challenges, we propose E-TransConvNet, a hybrid framework that integrates a CNN backbone with a single enhanced transformer block, effectively balancing computational efficiency, local continuity, and global contextual reasoning. The enhanced transformer block introduces two novel components: a Convolution Attention Block (CAB), which captures global context while simultaneously modeling local spatial relationships by integrating depth-wise convolutional operations into multi-head self-attention mechanisms, and a Progressive Feature Enhancement Feed-Forward Network (PFE-FFN), which further refines boundary delineation by selectively capturing features across various receptive fields to preserve fine-grained structural details. Extensive experiments are conducted on six public medical image segmentation datasets spanning two medical imaging modalities to validate the capacity and effectiveness of E-TransConvNet. Our model achieves superior segmentation accuracy and enhanced boundary delineation across all evaluated datasets. Specifically, E-TransConvNet outperformed MedSAM by an average of 3.55 % in terms of Intersection over Union (IoU) across the six datasets. The source code for E-TransConvNet will be available at https://github.com/S-domain/E-TransConvNet.