Study and Design of Deep Learning Computer-Aided Diagnosis Systems Based on Biomedical Images and Signals

This Ph.D. thesis aims to describe the research works conducted for the design, the development and the evaluation of innovative Computer-Aided Diagnosis (CAD) systems based on machine learning and deep learning techniques. Several CAD solutions were developed in different medical applications trying to ensure, when possible, three main CAD requirements: improve clinicians performance, reduce or at least not increase clinicians time and integrate the CAD solution in standard procedures. The proposed applications involved images and signals processing; the firsts required the use of different deep learning models to face classification, detection and segmentation problems, while the latter allowed to investigate machine learning as signal processing technique for movement disorder analysis and for a more speculative research in the rehabilitation field. In order to properly validate the proposed algorithms, all the methodologies were applied on real data provided by clinicians, public datasets or specific acquisitions. Potentialities, challenges and drawbacks about deep learning for medical imaging analysis are discussed in two medical fields, digital pathology and radiology, and complete pipelines are proposed to accomplish three clinical practices: global glomerulosclerosis analysis for Chronic Kidney Disease evaluation, kidneys volume analysis for Autosomal Dominant Polycystic Kidney Disease evaluation and organs segmentation for generic volume quantification. Each study case aims to identify and overcome the limitation of classical image processing techniques, and paves the way towards the clinical use of CAD systems based on deep learning. A second part of this thesis focuses on machine learning and deep learning for signals processing; deep neural networks were investigated for movement disorders analysis and a particular neural model for surface electromyography analysis has been proposed for the evaluation of complex muscle activation patterns, useful in the rehabilitation field. The developed solutions for signals and images processing, were compared with literature standards and, if possible, a personalised classical pipelines has been proposed and customised to face each clinical challenge. The thesis is divided into six chapters. The first chapter provides an introduction about the reference context. The following chapter two describes the state of the art about traditional CAD systems based on conventional machine learning algorithms, and the novelty that deep learning techniques bring to CADs and medical practices; description of the main convolutional neural network models and autoencoders, and literature about the application of deep learning and machine learning to the concerned medical fields are reported. Chapters three, four and five report the original contribution about the application of deep learning and machine learning techniques to the two types of medical data: images and signals; in detail, chapter three reports the applications in the clinical areas of digital pathology and radiology, focusing on the development of full pipelines based on image analysis; chapter four shows a more speculative research work for signal processing, focusing on the application of undercomplete autoencoders for surface electromyography analysis; chapter five reports the applications of deep neural networks for diseases assessment and grading in subjects affected by movement disorders. The analysed study cases and the contributions reported in this thesis were compared with standard processing techniques ad-hoc developed. Finally, the conclusions about the research works and proposals for future researches are reported in chapter six.