Software Quality Assurance per Sistemi Medici

Medical and health processes are increasingly dependent on software that plays a critical safety role in diagnosis, decisions, and device control or functioning. Software is present in many applications: from the one embedded in medical systems to the application installed on PCs for patient management, from protocols that allow communication between different medical devices to software that assists in diagnosis. The health and safety of patients (and sometimes doctors) who interact with medical devices depend on the correct functioning of each medical device and, in particular, on the software running on them. This is the reason why, in recent years, researchers and industries have focused on creating a well-defined software life cycle for medical software and on verifying, validating, and testing medical devices, which must be mandatory activities during development, as required by the certification standards. This thesis presents several attempts to define methodologies and strategies that may be used during the development of medical devices. In particular, model-based testing has proven to be a valid technique that allows easier compliance with the certification standards, since using formal methods for modeling the system under analysis allows obtaining in a more rapid and effective way documents and testing the system in a more satisfactory way. As suggested by the literature, for complex systems, the technique of Combinatorial Interaction Testing has proved to be suitable for reducing the effort of testing, without losing effectiveness. In this thesis, these methods have been applied to several case studies, namely the MVM (a mechanical ventilator that I contributed to develop and test during the initial phase of the COVID-19 pandemic), the Pill-Box, and the PHD protocol. Recent software and medical devices also include many components that use artificial intelligence (AI) to process images and data and produce results. Model-based or combinatorial techniques are not always suitable to deal with AI, and so other methods, mainly based on the robustness analysis of neural networks, must be used in order to validate AI-based medical software. This thesis presents the concept of robustness w.r.t. external input perturbations, both for neural network classifiers and estimators and applies them to real case studies, namely to a classifier for Breast Cancer diagnosis and an estimator of the pO2 level in blood.