Publications

At this page we provide you with an overview of our methodology developing AI-based solutions and present a selection of the most relevant publications.

In a recent review article in MED Engineering we summarized our development and validation process of AI-solutions in radiology and provided an outlook on future developments.

The full text is available here (German only): MEDeng 3 2020

Based on repetitive scans of the same patients with Alzheimer’s disease and healthy controls from the MIRIAD-study taken at the same day mdbrain was validated against the academic gold standard Freesurfer (Harvard Medical School). The deviation from the mean of repetitive scans was in the range of the expected physiological variance over 1 year and significantly better compared to Freesurfer.

These results were presented first at ECR 2020. A full publication is in preparation.

Based on scans of patients with Alzheimer’s disease and healthy controls from the MIRIAD-study the power of mdbrain vs. the academic gold standard Freesurfer (Harvard Medical School) was analysed with respect to their ability to differentiate healthy from diseased patients. Taken volumetric results of Hippocampus and temporal lobe together in this cohort a sensitivity of 1.0 and a specificity of 0.91 was reached by mdbrain. There was a trend toward superiority of mdbrain over freesurfer, but the difference was not statistically significant.

These results were presented first at ECR 2020. A full publication is in preparation.

Based on examinations of 24 patients with histologically confirmed epileptogenic hippocampal sclerosis, the precision of mdbrain’s volumetry solution was investigated. Atrophy was correctly detected with a specificity of 100% and a sensitivity of 96%, while the BrainMorphometry solution from the Siemens Syngovia suite used for comparison had a sensitivity of only 36%.

These results were first presented by Anna-Lena Mayer (University Hospital Erlangen) at the neuroRAD Congress 2021. A full publication is in preparation.

Based on blinded examinations of 35 patients with clinically confirmed diagnosis of dementia (frontotemporal or Alzheimer’s), as well as 20 healthy controls, the extent to which mdbrain‘s volumetry solution  influences the diagnostic precision was investigated. In general there was a gain in diagnostic precision and by the additional pair of eyes provided by the AI solution, whereby the effect was stronger the lower the level of radiological training was.

These results were presented for the first time by Jan Rudolph (LMU Munich) at the neuroRAD congress 2021.

Based on 3D-MRI-scans of Multiple Sclerosis patients from the LITMS-Study we have compared the performance of mdbrain against neuroradiological consensus reading as ground truth and two academical standard tools (LST and SLS). In summary we could show, that mdbrain performed better than LST or SLS in particular with respect to smaller lesions translating into a better F1-Score.

These results were presented first at EAN 2020. A full publication is in preparation.

Note: This study was done with an older version of mdbrain. Newer versions include improved lesion algorithms.

Based on blinded studies of 101 patients with confirmed diagnosis of multiple sclerosis (MS) according to McDonald criteria, the extent to which mdbrain‘s lesion module correctly detects and regionally classifies inflammatory lesions was investigated. The expert judgment of three neuroradiologists was used as the gold standard. At the same time, it was investigated whether the methodological step towards fully DeepLearning-based algorithms was beneficial for the precision of lesion detection. The results showed significant improvements, particularly in sensitivity, in all regions, with potential for improvement in the area of infratentorial lesions.

These results were first presented by Stefan Hock (Erlangen University Hospital) at the neuroRAD Congress 2021.

Note: The current version of mdbrain has been further improved with respect to the quality of lesion characterization since this work!

Semiquantitative Scores like MTA, Erica or MtPA are widely used in radiology while lacking fast and automated methods to quantify brain regions of interest. Taking PSP-diagnostics (Progressive Supranuclear Palsy) as example the percentile based volumetry of the midbrain by mdbrain was compared to classical MtPR-Scores. Using our volumetric approach we could achieve even higher AUC values – although not statistically significant. These results were generated fully automatically in less than 5 min from MPRAGE-images.

These results were presented first at ECR 2021. A full publication is planned with more data included.

In radiological centres we often see the prejudice that running 3D-T1-Sequences as part of the standard protocol takes too much time and hence quantitative brain volumetry is not possible in daily routine. A workgroup from the TU Munich / Germany countered that myth. Brain MRI scans of the same patients were analysed by mdbrain and Freesurfer using different compression factors using the compressed sensing (CS) function of Philips-Scannern. Up to a CS-factor of 8 the values generated by mdbrain seemed to be quite robust corresponding to image acquisition times of less than 3 min…

Link to full publication HERE.

Newsletter
abonnieren!