IMPROVING THE QUALITY OF BIOMEDICAL IMAGES WITH COMPUTER VISION: DEEP LEARNING SYSTEMS WITH A HIGH LEVEL OF ABSTRACTION AND UNDERSTANDING FOR THE RECOGNITION AND CORRECTION OF MRI ARTEFACTS
When we speak of biomedical imaging techniques, we are referring to that series of physical phenomena that are exploited to generate images of the internal tissues of the human body or parts of it that are not visible from the outside, in a totally non-invasive way for the patient. The Magnetic Resonance Imaging (MRI), Computer Tomography (CT) and Positron Emission Tomography (PET) techniques are mainly included within these technologies.
Medical imaging is one of the fields of applied research where Deep Learning is making significant contributions to support specialists in the field. However, its actual implementation is taking place very slowly due to the lack of large amounts of truly usable data. There is, in fact, a high degree of heterogeneity of sources, a poor level of annotation and harmonisation of data as well as privacy problems that often prevent them from being shared between healthcare facilities and research groups.
The automated tools that AI is able to develop include those for disease detection, quantification of the extent of a disease, disease characterisation and numerous other decision-supporting software tools for the physician. In the field of medical image processing, the main activities are detection, classification, segmentation, enhancement, reconstruction and registration.
In this area, SCAI Lab’s R&D team has recently been developing a series of specific algorithms that exploit the latest DL (Convolutional Neural Networks) techniques used in the field of Computer Vision. The aim: to identify, localise, classify and finally significantly reduce the presence of artefacts on MRI , i.e. to minimise visual disturbances that may occur on diagnostic images.
During MRI scans, there are several reasons why artefacts can occur on the resulting image: problems with the machine’s hardware and/or software, involuntary movements of the patient or the presence of prostheses with metal components. In these cases, the image presents an artificial feature that is not really present in the object under investigation and can seriously compromise the diagnostic work in some cases, in others even simulating the onset of specific pathologies.
Detecting and correcting artefacts is of essential strategic importance, for example, in the context of pre-operative planning and post-operative follow-up of a prosthesis implant (whose metal components are a primary source of disturbance on MR) or in the identification of a chronic pathology with neurological pathologies that prevent the patient from remaining still during scanning. The challenge is to propose highly efficient methods to improve the quality of imaging and, more generally, of the doctors’ clinical-diagnostic systems, in an increasingly patient-centred perspective. Systems that enable the attainment of more reliable and accurate results at all stages of the care pathway, to drastically reduce the invasiveness of tests, to improve the quality of patient care and treatment while containing the relative social and management costs.