IDUNet-ERNN: Generic Biomedical Image Segmentation and Annotation Using Deep Learning Model

Annotation and segmentation systems for biomedical images are essential to the process of segmenting images to identify specific diseases. However, training accurate segmentation models requires a substantial amount of precisely annotated data, which is time-consuming, labor-intensive, and prone to human error. This limitation hinders the development of efficient and reliable automated disease detection systems. Therefore, this work introduces a novel biomedical image segmentation and annotation technique. Initially, pre-processing the biomedical image for denoising and contrast enhancement is done using the Gaussian non-local mean filter and the dynamic histogram equalization algorithm (GNM-DHE). Then, the improved dense U-Net with enhanced region-convolutional neural network (IDUNet-ERNN) model is utilized for segmentation and annotation. RNN uses a region proposal network (RPN) to improve performance, and IDUNet is enhanced with a triple-dilated module. This new method produces effective segmentation and annotation results for biomedical images. Finally, chaotic-based Aquila optimization algorithm (C-AOA) is utilized to reduce the loss function. Then, the overall performance of the proposed model is analyzed using the biomedical image dataset named the Medical Segmentation Decathlon (MSD) dataset, which includes nearly 10 tasks or organs. To determine whether the proposed model is superior or not, its performance is contrasted with that of many related strategies. The accuracy of the proposed model is obtained to be 98.27%, and it also achieves 98.27% sensitivity. The proposed technique obtains 0.97 mean intersection over union (MIoU), and several performance metrics are analyzed to determine the efficiency of the proposed segmentation model. Then, it relates the achieved performance of the proposed technique to some other related techniques.