Emergence of deep learning (DL) and its applicability motivated researchers and scientists to explore its applications in their fields of expertise. In medical technology, a huge amount of data is required, and dealing with huge data is a challenging task for researchers. The emergence of neural networks and its modifications like convolutional neural networks (CNN), generative adversarial network (AGN), recurrent neural networks (RNN), and their subcategories has provided a stage to flourish deep learning. DL has been a successful tool in the fields of pattern recognition, natural language processing (NLP), image processing, speech recognition, computer vision, etc. All these techniques have been employed in healthcare. Image processing has been proven to be a fruitful technique for physicians to properly diagnose patients through CT scan, MRI, PET, radiography, nuclear medicine, ultrasound, etc. In this chapter, some applications of DL in healthcare have been envisaged, and it has been concluded that this technique is very successful in healthcare.
Top1. Introduction
Deep learning (DL), a powerful technology for dealing with large datasets, is a branch of a popular technology known as machine learning. The relation between deep learning, machine learning (ML) and artificial intelligence (AI) is depicted in figure1. Artificial neural networks (ANN) or simply christened as neural networks (NN) are the backbone of DL. The basic difference between the neural network and deep learning is that neural networks consist of three layers: first one the input layer, second one the hidden layer and third one the output layer while deep learning has more than three layers. In deep learning, there is more than one hidden layers (Bengio 2009, Schmidhuber 2015, Kumar, et al., 2016, 2017; Tripathi, et al. 2022; Dhar, et al., 2023).
Depth is the factor which distinguishes the DL networks from the neural network which has only one hidden layer. In DL networks, data passes through more than one hidden layer in pattern recognition in multistep process. This enables DL networks to have higher learning capabilities with more precision. DL networks are very efficient in dealing with the huge amount of data. Hence, they can solve the complex problems with pace and efficiency (Pan and Yang, 2010). The drawback of DL networks is that they are not applicable on small datasets.
Figure 1. Comparison between ANN and deep architecture
With the emergence of computers and medical technology, the last four decades have witnessed great deals of biomedical data which are available in abundance. But it is a difficult task to deal effectively with the biomedical data as these are noisy, sparse, chaotic and with high dimensions. Therefore, to acquire efficient information from biomedical data, researchers were in high need of a technique which could overcome this problem. The emergence of deep learning has shown a ray of hope for researchers. Deep Learning is an efficient technique for speech recognition (Abdel-Hamid et al., 2014), Computer vision (He et al., 2016), NLP (Lan et al., 2020), etc. DL has been extensively applied in several fields of healthcare like drug development, clinical image processing, disease detection, electronic health record, DNA sequencing, etc. The beauty of Deep learning is that it automatically extracts the features from the data and plays a vital role in sparing time and resources.
Top2. Common Deep Learning Algorithms
Numerous DL algorithms have achieved popularity in different fields. But the algorithms popular in healthcare are autoencoders, CNN, RNN and belief networks. We will discuss these here in brief.
2.1 Convolutional Neural Networks (CNN)
One of the prominent DL algorithms, CNN has extensively been employed in several fields like financial time series, natural language processing, image classification, medical image analysis, video and image recognition, image classification, recommender systems, image segmentation, brain–computer interfaces, face recognition (Rai et, al. 2022), speech recognition, etc. CNNs are a specific type of ANNs in which the mathematical operation convolution is used rather than matrix multiplication in one of their layers. CNN structure involves three layers: first one is termed as the convolutional layer; second one is termed as the pooling layer and the third one is termed as the fully connected layer. Convolutional layers perform the role of feature extractor, pooling layers perform the role of reducer of dimensions of data whereas fully connected layers perform the role of classifiers (Tripathi, et al. 2022).