Published in
Clinical Neuroradiology, 31(Supplement 1), 41-42. (2021)
Authors
S. W. Hock, D. C. Marterstock, A.-L- Mayer, C. Bettray, K. Huhn, V. Rothhammer, A. Dörfler, M. Schmidt
Abstract
Purpose
Artificial intelligence (AI) algorithms have already had a major impact on medical imaging and opened a wide field of detection of textural and morphological patterns. Aim was to evaluate the potential of latest AI regarding diagnosis and follow-up of Multiple Sclerosis (MS) in clinical radiology.
Materials & Methods
We included patients who had undergone MRI at a single academic hospital. MS lesions were evaluated according to McDonald criteria by latest and previous generation AI (Figure 1) and three expert neuroradiologists (gold standard) independently. Following statistical metrics were calculated and compared: Sensitivity (TPR), specificity (TNR), overall accuracy (ACC), false positive rate (FPR) and Dice similarity score (DSC).
Results
A comparison of ANN corroborates the superiority of the latest generation AI compared to the previous generation in detection of MS lesions (Figure 2). Overall sensitivity (77% vs. 29%) and DSC (0.81 vs 0.39) of the latest version AI were significantly higher. In the periventricular compartment TPR (77% vs. 53%), ACC (92% vs. 87%) and DSC (0.8 vs 0.64) were higher, while TNR (96% vs. 96%) and FPR (0.043 vs. 0.041) did not change significantly. In the juxtacortical compartment TPR (62% vs. 0.5%), ACC (95% vs. 90%), FPR (0.018 vs. 0.001) and DSC (0.7 vs 0.01) were higher, while TNR (98% vs. 99%) did not change significantly. In the deep white matter TPR (82% vs. 46%), ACC (82% vs. 62%) and DSC (0.85 vs 0.6) were higher, while TNR (80% vs. 86%) was lower and FPR (0.14 vs. 0.20) did not change significantly. Infratentorial TPR (53% vs. 16%) and DSC (0.69 vs 0.27) were higher, while TNR (99% vs. 99%), ACC (97% vs. 95%) and FPR (0.015 vs. 0.015) did not change significantly.
Discussion
Preliminary data show that the latest generation AI provides consistent, automated, and fully reproducible assessment of MS lesions without being influenced by intra- and/or interobserver, intrinsic human variability – especially in the context of longitudinal patient follow-up. Thus, it may aid reliability and standardization in diagnosis and follow-up imaging of MS.
