ResBCU-Net: Deep learning approach for segmentation of skin images. (January 2022)

• Record Type: Journal Article
• Authors: Badshah, Noor; Ahmad, Asif
• Is Part Of: Biomedical signal processing and control. Volume 71(2022)Part A
• Publication Date: 2022-01
• Subjects: Image segmentation -- Neural network -- Convolution -- Pooling -- Batch normalization; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103137
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Designed a new DNN based on U-Net, Residual blocks, batch normalization and LSTM-Net. The designed Architecture is applied for segmentation of skin images. Tested on Skin images data set and achieved about 94% accuracy. Abstract: Large networks on big datasets have put great impact in success of deep neural networks (DNNs). The greatest breakthrough achieved in this field in 2012 by successful training of a large network, AlexNet, on ImageNet dataset containing 1 million images. U-Net architecture, a DNN based on Convolutional Neural Networks (CNNs), immensely advanced segmentation of medical images. In the present times, neural networks have outperformed other state-of-the-art approaches in segmentation of images. In this paper, we present a neural network based on the CNNs for segmentation of medical images. The network, ResBCU-Net, is an extension of the U-Net which utilizes Residual blocks, Batch normalization and Bi-directional ConvLSTM. In addition, we present an extended form of ResBCU-Net, ResBCU-Net(d = 3), which takes advantage of densely connected layers in its bottleneck section. The proposed neural network is trained and evaluated on ISIC 2018 dataset, which is publicly available dataset containing 2594 melanoma cancerous skin images. The network inferences segmentation of the images more accurately than other state-of-the-art alternatives.