AI in medical imaging: from proof-of-concept to daily radiology services

Despite the significant amount of research in the area of AI for medical image applications, adoption of these new tools in the radiological practice is limited and potential benefits for the patients are therefore not yet achieved or proven. This project will provide unique solutions to successfully integrate AI in the radiological practice. In this project we will start from radiology driven requests, search or develop candidate AI algorithms and work out standardized, generic procedures up to the successful, safe implementation of the new tools. In this multiphasic project AI algorithms will be made compliant with current legal and ethical standards related to AI medical software, which involves, among others, ethics, General Data Protection Regulation (GDPR) and Medical Device Regulation (MDR). Furthermore, it requires automated testing with optimized databases and adjustments by the medical team. The next phase solves integration in the routine workflow, from the automated and verified selection of the proper image series up to the patient specific use of the results in a structured report. The value of the software can be expressed in terms of quality of life (patients) or the more efficient use of the means (personnel and devices). Applications applied on a hospital wide scale could generate big data that, if managed properly, could be used for population statistics and possibly predict new pathologic conditions in an early stage. In parallel to this project, the safe and best use of the tools has to be checked with a test protocol inspired by medical physics testing, as medical physicists have the legal task to approve devices and also software that may impact the quality of medical images. Also, it has been shown that acceptance of new AI solution can be problematic and appropriate measures need to be installed to maximize the trust of the radiologist in the AI solution. In a first case study, we will focus on the radiological request for kidney segmentation of anatomically normal kidneys and subsequently for the special situation of patients suffering from cancer, and then Autosomal Dominant Polycystic Kidney Disease. The tool should start from applications on selected (pre-transplant kidney) patients up to the creation of a large data base of normals that can be used for outlier tracking. A second case study has still to be determined, but may be of another type such as prioritization of x-ray exams pending the detection of specific diseases. For this project to be successful new research at the PhD level is needed to efficiently implement compliance with legal, GDPR and MDR requirements, to develop new clinically acceptable test methods for AI algorithm approval and to work out new data base structures . The final paper shows the health gain for the patients and discusses how to streamline this process for the safe introduction of more AI algorithms.