Cranial MRI (cMRI) is a standard procedure for the diagnosis of neurological or psychiatric disorders. Compared to cranial CT scans, cMRI imaging provides a wealth of additional information and is also free of potentially harmful X-rays. With the help of many sequences of different weightings, cMRI can display different information about the brain tissue. However, as these sequences are time-consuming, not all available sequences are simply performed; instead, the examination programmes are compiled in such a way that the specific question – for example, regarding ischaemia, tumours, inflammation or dementia – can be answered. A common incidental finding in cMRT is intracranial aneurysms, which affect about 2-3% of the population. These are bulges in the arteries supplying the brain, which are usually asymptomatic but, if they rupture, can lead to a specific form of cerebral haemorrhage associated with high mortality and morbidity.
Aneurysms are rarely detected in standard MRI sequences, but mainly in special vascular imaging techniques such as time-of-flight angiography (TOF), which reads the signal of the flowing blood itself and therefore works without contrast agents. Since this also takes between 5 and 12 minutes, depending on the desired quality, it is not routinely performed on all patients, but only in cases of vascular issues. As a result, many aneurysms are simply not detected in cMRT examinations without TOF sequences. One basic sequence that is performed in every cMRT examination, however, is T2 weighting. The idea behind CONAD is to automatically perform an additional AI-based search for any existing and previously undetected aneurysms in these T2 sequences during EVERY MRI examination. If the AI finds a possible aneurysm, it reports the suspicion to the radiologist, who can then expertly examine the MRI image and, if necessary, clarify the situation further with an additional TOF angiography. With timely diagnosis, the aneurysm can be eliminated before it ruptures, and the patient has a significantly higher life expectancy. From a medical and preventive perspective, it therefore makes perfect sense to detect existing aneurysms as early as possible.
From an AI perspective, the challenge lies in detection in the T2 sequences without the additional vascular contrast provided by an angiography sequence. Contrast differences between vessels and tissue are very small and often obscured by noise. Pre-processing with suitable image processing algorithms or style transfer approaches is therefore essential. Another issue is the anisotropic resolution of MRI scans, which is usually four times greater in the z-direction than in the x- and y-directions. Specially adapted 3D deep learning methods must therefore be developed and used to detect aneurysms in MRI scans.
Partners
Universität des Saarlandes, Lehrstuhl für Neuroradiologie
