MedSegNet: A Lightweight Convolutional Network Combining Dual Self-Attention and Multi-Scale Attention for Medical Image Segmentation

In this work, we propose a novel lightweight convolutional neural network called MedSegNet that innovatively incorporates residual modules with a fusion of dual self-attention and multi-scale attention mechanisms, designed for the segmentation of three different types of medical images such as CT, non-mydriatic 3CCD, and colonoscopy images. This network has demonstrated proficiency in executing segmentation tasks for images of a specific modality, depending upon adequate training with a representative dataset of that same modality. Experiments are implemented to train MedSegNet using three different datasets, and the trained MedSegNet is tested on the respective test datasets. In comparative analyses with state-of-the-art models, MedSegNet has been shown to outperform in terms of segmentation dice coefficient (DSC) and intersection over union (IoU), computational efficiency, and robustness to variations in medical imaging modalities. These results highlight the potential of MedSegNet to set a new benchmark for medical image segmentation tasks.