ABSTRACT
In recent days, systems are designed to have better classification based on input. The inputs vary based on the application that is intended, viz., retina image for detection of diabetic retinopathy (DR). Statistical data for classification based on various input parameters, but these are not limited. This chapter mainly focuses on the issues of biomedical image processing, having its roots in deep learning (DL). Existing techniques before the evolution of DL have made their mark, but its performance is limited. These techniques fail as the database (DB) size increases. In addition, various constraints need to be considered while calculating the recognition accuracy. Therefore, it is the need of the hour to find feasible solutions for improving the classification accuracy. Primarily these techniques are classified as supervised and unsupervised classification techniques. It is a well-known fact that supervised techniques have apriori data while the latter does not. Obviously, the former results in better accuracy, and the latter does not. In the real world scenario, unsupervised classification is more relevant as the input data tends to vary with time.
DL is the fastest-growing field that has already proven its worth in machine learning. The attempt to model large-scale data by the use of various layered networks led to the development of deep neural networks (DNN), providing an application to various fields, but not limited to viz., image classification, 282and pattern recognition. In general, DL has two properties, one is various layers that support non-linear processing, and the other is supervised or unsupervised learning based on the features existing on each of the layers. Before DL development of artificial neural networks (ANNs) has enthralled into science and technology, and from 2006, DL created its impact, and now it’s so deep that still, scientists are learning it. An essential part of DL marks its path into optimization, a task that provides data that best fits transfer function and finds a better fitting curve. Such optimization has various applications in classification and recognition. As in the case of an Iris recognition system, as the DB size increases, the features representing a particular class of Iris must also increase, or else the system tends to fail. Hence, optimization is required to capture optimal features for proper recognition.
Another application of DL in classification and recognition is in genomic signal processing. The classification of a large set of genome data is a trivial task, and keeping in mind the classification of such an extensive data set is hectic. The systems so trained and developed must be capable enough to handle such extensive data. DL is one of the applications that help users to find a particular genome code resulting in the identification of a particular disease or ailment from the given more extensive data set. The system tends to become more complicated for the search of a particular genome sequence, given different sequences of particular organism viz., COVID-19.
This chapter discusses DL architectures that are used in biomedical image analysis viz., CNN, DBN, and RNN. Architecture is SAE, which has found its application in skin analysis, detection of organs in 4D patient information, segmentation of hippocampus from infant’s brains, optic disc extraction from fundus images. DBN is another architecture that has provided various applications such as segmentation of the left ventricle from heart MRI images, identification, and segregation of various retinal diseases. DBN, DNN, and RNN are different DL architectures that are used in the analysis of the genome sequences viz., finding the splicing patterns in individual tissues, highlighting the pathogenicity of genetic variants, identification of splice junction at DNA level, and understanding of the non-coding genome, identification of miRNA precursor and targets. Protein structure is also predicted using DBN, CNN, and RNN. Various models were developed for identification of structural binding choices, and prediction of binding sites of RBPs, disordered protein, other structures, local backbone positions, the area covered by proteins.
Keeping aside the advantages and uses of DL, problems while applying DL algorithms in biomedical applications persist. Besides having achievable 283accuracy and better speed, the computational complexity of these algorithms also increases based on classification methods. Labeling of medical images is a hectic task that requires professional training. Instead, these images have a privacy lock that cannot be used by the general public and researchers. Therefore, getting such data is also a complicated process.
Further, getting such a large amount of data is also a trivial task. Still, now metrics for the classification process are under development. The developed metric must also be uniform in assessment for various types of data and techniques that are existing in the networks. With the availability of such large data sets, it is also necessary to analyze the developed model carefully and adapt the model based on the features, properties, or characteristics of the data. As technology is accessing more data, which is accessed by wearable sensors via smartphones, DL helps as a tool for interpreting such data, detection, prognosis, prevention, diagnosis, and therapy.
