Applications of AI and Deep Learning in Biomedicine and Healthcare

Deep Learning And Its Use In Medicine

Medicine and biology are becoming more and more data-intensive fields. These data's bulk and complexity offer new possibilities as well as difficulties. In privacy-sensitive settings, automated algorithms that identify significant patterns may produce knowledge that is useful and transform the way we classify patients, create treatments, and research diseases.

The phrase “deep learning” now refers to a group of novel methods that collectively have shown revolutionary improvements over state-of-the-art machine learning algorithms in a number of domains. For instance, because of their great accuracy and versatility, these techniques have transformed speech recognition and image categorization over the last five years. Deep learning algorithms have recently showed promise in a variety of domains, including dermatology, computational chemistry, high-energy physics, and written language translation.(Jain,1996) These algorithms' “off-the-shelf” implementations have yielded accuracy that is on par with or better than that of earlier best-in-class techniques that needed years of intensive customisation, and specialised versions are currently in use at industrial sizes.

Numerous functions of deep learning are similar to those of other well-known ML(machine learning) techniques. Deep learning techniques may be utilised as a type of clustering in both supervised and unsupervised, or “exploratory,” applications. In applications which are supervised, the goal is to predict accurately many labels associated with individual data point.(Pour, 2017) It is possible for deep learning techniques to combine both of these phases. Specifically, these techniques generate features specific to a given problem and combine them into a prediction once sufficient labelled data are available. Similarly, one may consider a single-layer neural network with continuous outputs to be analogous to linear regression. Hence, supervised deep learning approaches can be regarded of as a more flexible form of regression models, particularly when it comes to mimicking nonlinear correlations between the input characteristics.(Norgeot,2019) These deep learning approaches are now superior to traditional machine learning methods for a wide range of problems thanks to recent advancements in hardware and massive training datasets.

The recent boom in this field of study is made possible by several significant advancements. The field's methods are now accessible to a larger population of computational scientists thanks to user-friendly software packages. Moreover, additional methods for quick training have made it possible to use them with bigger datasets. Networks become more resilient when nodes, edges, and layers drop out, even when there are a lot of parameters.(Mamoshina,2016) Lastly, the larger datasets that are currently accessible are enough for fitting the wide range of deep neural network parameters. Deep learning is very adaptive and can manage the small fluctuations of each domain to which it is applied because these parameters are currently convergent.

Updated research related to biomedical field is covered in this chapter, and the most effective applications choose neural network designs that are appropriate for the given task. In figure 1, we illustrate a few basic example architectures. Convolutional neural networks (CNNs) can emphasise local correlations in data that have a natural adjacency pattern. This is especially helpful if in the neural network's first stages convolutional layers are used. Because they don't need labels, other neural network topologies like autoencoders are frequently used for unsupervised tasks. A selection of example applications are shown in Table 1, however the neural network architectures have all been applied extensively to a wide range of biological data sources.(Ching,2018)