1 - Congruence of deep learning in biomedical engineering: future prospects and challenges

Abstract

Abstract Deep learning models have opened up many prospects in medical images for achieving unprecedented performance, for example, classification of tissues and division or segmentation are a few medical outcomes. This chapter evaluates and describes the convolutional neural network (CNN) intended for characterization of tissue in clinical imaging, which is applied for segregating essential metastatic liver tumors from diffusion-weighted magnetic resonance imaging information. Advancement in the field of deep learning for normal pictures has provoked a surge of enthusiasm for applying comparative strategies to clinical images. Most of the initial attempts replaced the input of a deep CNN with medical images, which does not consider the basic contrasts between these two kinds of pictures. In particular, fine details are fundamental in clinical pictures, unlike regular images where coarse structures are very important. This distinction makes it difficult to utilize the current organized models created for common pictures, because they chip away at downscaled medical images to decrease the memory prerequisites. These subtleties are important to provide accurate detection. Furthermore, a medical test in clinical imaging regularly accompanies many perspectives, which must be intertwined to arrive at the right conclusion. A survey of deep learning is used for image classification, carotid ultrasound data investigations, cardiotocography, intravascular ultrasound reports, lung computed tomography reports, brain tumor prediction, coronavirus prediction (COVID-19) object detection, segmentation, breast cancer prediction, electrocardiogram signals, electroencephalograms, photoplethysmographic signal registration, psoriasis skin disease, as well as cancer detection. Concise summaries are delivered of trainings per application zone: pulmonary, musculoskeletal neuro, digital pathology, abdominal, retinal, breast, and cardiac. There are various types of deep learning techniques present to improve the accuracy of the medical dataset. Deep reinforcement learning, recursive neural network, multilayer perceptron, recurrent neural network, Boltzmann machine, and CNN are different types of deep learning techniques used to train the image and signal dataset. Generative adversarial network (GAN), autoencoder, and deep belief neural network are subcategories of unsupervised pretrained neural network. Some well-known architectural models of CNNs are ResNet (2015), VGGNet (2014), SqueezeNet (2016), GoogLeNet (2014), and ZFNet (2013) and are the visualization concept of the deconvolutional network; AlexNet (2012) and LeNet (Peng et\xa0al., 2009; Mitchell; Bengio, 2012; Dutkowski et\xa0al., 2015; Han et\xa0al., 2020 [ [16] , [17] , [18] , [19] , [20] ]) are basically used to train image datasets; the long short-term memory technique is used to train signalized datasets; and RHSBoost and genetically optimized neural network are used for efficient multiple classification of datasets. Dimensionality reduction, feature extraction, overfitting, underfitting, and normalization problems can be solved using various types of optimization algorithm. Image security is another important part, and by using an autoencoder, GAN network, and CNN we can prevent alteration in the medical image. Minor alteration of the medical image is very dangerous to patient life. By using deep learning and steganography, we can first compress as well as train the dataset, then security can be preserved after embedding of watermarks (which is a secret image visible to the human eye that cannot be altered; this steganography concept is called watermarking).