The growing development of deep learning and its increasingly widespread use signals a trend to convert to a new standard in each application area, consolidating itself as a general purpose technology. In the general analysis of innovations based on the use of deep learning on different applications of medical images, it can be inferred that its use is increasingly widespread and that it is driving greater innovations and transformations at an increasingly greater speed. Additionally, it is evident that the use of deep learning in medical imaging for the diagnosis of various anomalies incorporates a significant number of the elements and methodological and organizational aspects identified under this study (e.g., invention methods from the perspective of application of deep learning). Therefore, this study emphasizes various aspects of deep learning models to have opted as a ready reference while examining the medical images as a de-facto standard.
Top1. Introduction
Advances in AI in recent years make it possible to create new technologies and have more and more applications in various fields. Its foundations, complex algorithms, are intended to solve all kinds of intuitive and subjective problems that until now, computers were not capable of solving, and were the domain only of human beings, capable of solving them automatically i.e. recognizing someone in a photo, or the medical field, detect the presence of an abnormality in an image or a signal (Galaz et al., 2021).
To achieve good performance by computers in these problems, it is sought to make them learn from experience, to learn reality in the form of simple concepts that are related to each other forming more complex ones, so that there is a hierarchy of concepts, with many layers. From this learning model, with multiple levels of representation and abstraction, what is known as Deep Learning arises (Athanasios et al., 2018). Computers that work with Deep Learning algorithms have to be able by themselves to acquire their knowledge, extract their deductions, their patterns from the data provided. This learning capacity is what is known as Machine Learning. (Machine Learning), and encompasses the aforementioned Deep Learning (Linardatos et al., 2020).
However new this field may seem, Deep Learning has been around for quite some time. The lack of fame for these techniques is mainly due to two reasons: 1) the unavailability of large amounts of data to train the models, and 2) not having enough hardware (HW) or software (SW) in the computers. powerful as for the execution of the models. In recent years, both problems have been solved, and their multiple applications have motivated growing interest in this field. From image classification and structure segmentation to speech recognition or object detection, the possibilities are many. For this reason, today it is being used by technology companies, SW infrastructure companies, and for scientific applications, among others (Ching et al., 2018).
The infinite scientific and medical applications of Deep Learning algorithms i.e. for computer-aided diagnosis, for making predictions from large amounts of data, for processing images in medicine, for designing drugs, or for constructing 3D maps of the brain are precisely the motivation for this work, as they can help improve patient health care, in terms of both precision and speed. A specific medical task where these algorithms are achieving very good results is the classification of lesions and tumors, that is, given a certain image, and based on its characteristics, assign the image one of the two possible output classes (i.e. malignant or benign tumor). In particular, helping the specialist to make better decisions or to look at certain parts of the image where the anomalies may be (Ker et al., 2017).
For this reason, because it is a very widespread application, and also because various tumors are spreading most frequently worldwide, in this study we want to pay special attention to Deep Learning techniques that can be used for diagnosis, monitoring, and evolution of tumors. Furthermore, after studying the state of the art in depth, a methodology based on a Deep Learning algorithm will be designed, the purpose of which is to classify these tumors.