A Lightweight Deep Learning Architecture for Efficient Multimodal Medical Image Segmentation Using Attention Mechanism

Medical imaging is essential for clinical diagnosis, prognosis assessment, therapy planning, and surgery.Deep learning architectures have made a significant impact in the field of medical image segmentation in recent years.U-net, SegNet, and Attention UNet are some of the well-known deep neural networks in this area and have been widely recommended architectures for medical imaging for a long time.Despite having exceptional outcomes in segmenting multimodal medical images, traditional U-Net design appears to be inefficient in some areas i.e. difficult datasets of various resolutions.The existing segmentation networks such as U-Net and SegNet are also prone to overfitting and contain semantic gaps within the encoder and decoder layer.Because of this, several changes to the already cutting-edge U-Net models are suggested.The goal of this paper is to provide an architecture that can effectively improve the traditional methods and process multimodal biomedical image datasets, including MRI, microscope, dermoscopy, CT scan, chest x-ray, and endoscopy images.An average IOU of 0.88 and a mean dice score of 0.93 for the six datasets was obtained.Compared to standard U-Net, there are fewer parameters and a shorter training time.There are just 28.34% of U-Net's parameters in the architecture.The architecture requires about 3.798M parameters and an average of 80 epochs to reach convergence.The architecture has also shown significant performance on various datasets.