EU-net: An automated CNN based ebola U-net model for efficient medical image segmentation

Abstract

Medical image segmentation (MIS) plays an important role in rapid disease detection and prognosis. U-Net, a widely used neural network in the field of MIS, encounters performance limitations when applied to complex datasets. These limitations are primarily due to the basic feature extraction blocks, namely the encoder and decoder, as well as the existence of a semantic gap between these two components. Some U-Net variants, such as Recurrent Residual U-Net, have attempted to solve the problem of simple feature extraction blocks by increasing the network depth. However, this approach does not effectively solve the semantic gap problem. In contrast, another variant, UNET +  + , addresses the semantic gap problem by introducing dense skip connections but retains the simplicity of its feature extraction blocks. To overcome these challenges, a new approach is required. Therefore, this research proposed an optimized weight and loss function network called CNN-based Ebola U-Net (EU-Net) architecture for MIS. In the first step, the images are pre-processed for noise reduction using a filtering method called bilateral filtering, which is a hybridization of two Gaussian filters. The CNN-based EU-Net acts as both an encoder and a decoder with convolutional layers. A modification in the skip connection is encouraged to minimize the semantic gaps between the encoder and decoder. In each encoding stage, the edge information of the input images is extracted using the Pyramid Edge Extraction Module (PEEM). Using the extracted information, a segmentation map is generated on the decoder side. The Ebola Optimization Algorithm (EOA) is used to reduce the loss function and overall weight of the proposed network model. To evaluate the performance level of the proposed network, four different datasets are used. The proposed model achieves an overall accuracy of 97.33%, an RMSE of 1.36 and a Dice coefficient of 94.97%.