Improved Automated Central Vein Sign Assessment by Multi-Level Classification
Till Huelnhagen, Omar al Louzi, Lynn Daboul, Jonas Richiardi, Daniel S. Reich, Tobias Kober, Pascal Sati · Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2023
Central vein sign (CVS) assessment has shown potential to improve differential diagnosis in multiple sclerosis, but automating this task remains non-trivial. As human inter-rater agreement was reported to improve by separating the tasks of lesion exclusion and CVS assessment, we hypothesized that this could also benefit automated CVS assessment. To test this hypothesis, we implemented a novel multi-level classifier for automated CVS assessment and trained and evaluated it in more than 9400 expert-reviewed lesions. The new approach outperforms previous methods, achieving per-class accuracies of 76%–83% in an unseen testing set and >90% accuracy to identify MS cases.