Deep Learning for data analysis on specific contexts (Automotive, Medical Imaging)

In light of the tremendous success gained by Deep Learning algorithms, the role of these techniques is becoming increasingly important in the challenging automotive and healthcare fields. The proposed dissertation collects all the research works done by the Ph.D. candidate, focusing on the development of advanced Deep Learning approaches for the mentioned domains. The contribution of this dissertation is not only to deliver effective Deep Learning solutions for a wide range of problems but also to outline the challenges encountered in these fields. Broadly, the scientific community of both areas has always faced the lack of high-quality datasets, which has affected the performance of DL algorithms as they directly depend on data used to perform predictions. In the context of the automotive industry, Deep Learning has opened new opportunities to process large amounts of complex data coming from multiple automotive-compliant sensors. In fact, current literature is focusing on the designing of advanced driving support functions to analyze the car driver's physiological status for assessing his/her fatigue level. However, collecting physiological data posed several challenges due to the expense in terms of time and resources, which has hampered the research investigations. Hence, the datasets used in such studies are often quite limited. Furthermore, the driving scenario has increasingly pressed the development of real-time DL applications in order to provide a faster response. Despite the effort to support this demand, run Deep Learning pipelines in real-time driving scenarios still remains an open issue. Motivated by these issues, in Chapter 2, we attempted to define effective solutions to overcome the limitations previously mentioned, copying with the adverse conditions. A common problem in the automotive and medical imaging domain is related to the studies assessment, which must follow strict protocols devoted to preserving the privacy of users along with their safety. This process has hindered the collection of high-quality datasets, especially in the medical imaging field. Medical imaging has always suffered by the limited availability of properly labeled data and the scarce quality data (without noise or artifacts). In addition, clinicians are always rather skeptical towards Deep Learning approaches, as deep networks do not often lead to understandable results because they operate as "black-box," which do not make interpretable the computations performed among their intermediate layers. In Ch. 3, we reported a fairly comprehensive discussion related to the mentioned drawbacks. Although we attempted to enhance the current literature by designing promising Deep Learning approaches, we also focus on treating the issues to be overcome by researchers, enlarging the discussion related to the main challenges encountered nowadays in this field. Finally, in the Appendices section, we reported our findings in the aerobiology and quantitative finance domains, where Deep Learning approaches have been applied for solving many complex tasks.