Development of artificial intelligence based systems for biomedical applications

The recent and more advanced applications of artificial intelligence (AI) reached a wide range of fields, transforming traditional workflows, perfecting current techniques, and introducing new paths not previously feasible. The common approach to AI-based systems relies on using an appropriately annotated database to train a model in order to identify a correlation between the input data and the desired output. With regard to the biomedical field, the literature does not provide a unified framework to follow for the creation of these tools, but rather proceeds heterogeneously. In this context, this work focuses on the study and creation of a framework intended to facilitate and enable the implementation of performing tools based on AI in the biomedical field. With the aim of providing an effective framework the production cycle of AI-based applications in the biomedical field has been studied. In particular, the framework was developed by analysing in detail all the implementation steps necessary for the development of new AI-based tools in the biomedical field and three main phases were identified: the clinical phase, the artificial intelligence engineering phase, and the application development phase. This work was carried out in collaboration with multiple medical research centers: the joint laboratory Custom3D, which brings together the Azienda Ospedaliera Universitaria Careggi and the Department of Industrial Engineering of the University of Florence; the T3Ddy laboratory, in collaboration between the Meyer Children's Hospital and the Department of Industrial Engineering of the University of Florence; a collaboration with the Center for Biomedical Technology of the Universidad Politecnica de Madrid. Within these partnerships, four case studies were analyzed to devise, use and refine the framework. The case studies concern the following medical sectors: 1) urology - with the study of renal tumors; 2) plastic surgery - for the automation of the production process of guides used for anatomical reconstruction of the ear; 3) psychiatry - for the identification of risk factors in patients with suicidal intentions; 4) neurology - for the evaluation of a therapy for the reduction and control of brain tumors. The implementation of the four case studies was carried out following the phases defined in the framework and using best practices for the implementation of artificial intelligence models. With regard to the first case study a model was developed to differentiate malignant clear cell renal cell carcinoma tumors and benign oncocytoma tumors, in the event that these are very small and difficult to interpret by expert doctors with a sensitivity of 94.59%. In the second case study, two AI-based tools were created to be used in the production process of surgical guides used by the surgeon to create anatomical replica of the patient's ear. In particular, these tools are able to generate the depth map from a simple image of the ear obtained from a normal camera, without the need to use more complex tools such as 3D acquisition scanners, with final MSE (mean square error) of ~0.07 and an average SSIM (structure similarity) of ~0.80, and to segment and identify the anatomical elements of interest within the depth map image of the patient's healthy ear with a 90% of accuracy considering each ear component as an independent class. For the third case, a classification model was developed using the clinical records of psychiatric patients. This tool is able to differentiate between two types of patients, those admitted for attempted suicide and those admitted for suicidal ideation with a final accuracy of ∼85%. Through the creation of this tool, it was also possible to carry out a study on the major risk factors that distinguish these two types of patients. Finally, for the last case study an application was developed that allows calculating the percentage of mouse brain volume occupied by glioblastoma multiforme tumors, reaching an average dice score of 84.48%. All for the purpose of evaluating the effects of optical hyperthermia to counteract and limit the growth and development of tumor cells through the use of nanoparticles of different materials.