A dual convolutional neural network with U-net and multi-residual network for medical image segmentation [EMBARGOED]

Abstract

Medical image segmentation is a growing field in medical image analysis. The segmentation of diseases, organs, or abnormalities in medical images has become demanding for the analysis of medical images. It is an essential task in medical image analysis for the identification of disease, treatment planning, and observation. Segmentation is a challenging job because of the numerous artifacts present in the medical images. Over the past few years, convolutional neural networks (CNNs) have shown outstanding performance in the segmentation of images. U-Net is a convolutional neural network, that has achieved excellent performance in the segmentation of medical images. Despite the outstanding performance of the U-Net and variants of U-Net, these architectures have certain limitations and are not sufficient for the segmentation of medical images. U-Net has an overfitting problem, a vanishing gradient issue, and it has a small receptive field. This thesis proposes an architecture, A Dual Convolutional Neural Network with U-Net and Multi-Residual Network for medical image segmentation to improve the segmentation. The proposed architecture consists of two networks. One network is U-Net, which uses pre-trained Resnet_50 as an encoder and the second network is MultiRes U-net. The residual block in the proposed architecture solves the vanishing gradient issue. Atrous spatial pyramid pooling (ASPP) block is used in both networks that increase the receptive field to capture more information. The proposed architecture is evaluated on three medical datasets: gastrointestinal polyp dataset, brain tumor dataset, and dental dataset. The gastrointestinal polyp dataset and brain tumor dataset are publicly available open-access dataset and is used by researchers for the segmentation. The dental dataset is collected from the Oral Radiology Unit, Kulliyyah of Dentistry, International Islamic University Malaysia. Data augmentation is applied to the training set to increase the dataset which will help to reduce the overfitting problem and improve the performance of the segmentation. The result of a proposed architecture, A Dual Convolutional Neural Network with U-Net and Multi-Residual Network, for medical image segmentation is compared to the existing architectures: FCN, U-Net, and Unet_Resnet. The performance of the proposed architecture is evaluated in terms of Dice Similarity Coefficient (DSC), Intersection over Union (IOU), Precision, and Recall. The experimental results show that the proposed architecture, A Dual Convolutional Neural Network with U-Net and Multi-Residual Network achieves higher DSC, IOU, Precision, and Recall for the segmentation of the gastrointestinal polyp dataset, dental dataset, and brain tumor dataset and outperforms the existing segmentation architectures.