Prof. Dr. Felix Eckstein (Freilassing; Salzburg / AT), Dr. Nicholas M. Brisson (Berlin), Susanne Maschek (Freilassing), Anna Wisser (Freilassing; Salzburg / AT), Prof. Dr. Francis Berenbaum (Lille / FR; Paris / FR), Prof. Dr.-Ing. Georg N. Duda (Berlin), Dr. Wolfgang Wirth (Freilassing; Salzburg / AT)
Abstract-Text (inkl. Referenzen und Bildunterschriften)
Introduction: MRI cartilage transverse relaxation time (T2) was shown to reflect matrix collagen and hydration composition, histological grading, mechanical properties, and early osteoarthritis status.
Purpose: To compare automated vs. manual cartilage segmentation for MRI T2 analysis in terms of a) segmentation agreement, b) cross-sectional and longitudinal T2 accuracy, c) sensitivity to between-group differences of knees that sustained anterior cruciate ligament (ACL) injury.
Methods: We studied 20 ACL-deficient patients with dynamic instability (non-copers; 162±91d post-injury), 22 ACL-deficient without instability (copers; 162±91d), 13 with surgical ACL reconstruction (104±65d), and 16 healthy controls. Sagittal MESE MRI (1.5T, TEx7: 9.7-68ms) was acquired at baseline (n=71) and 1-year later (n=55). Femorotibial cartilage plates were manually segmented by experienced readers, with quality control. The CNN algorithms were trained on MESE data from another study acquired with the same scanner (n=50). T2 was extracted from deep (50%) and superficial (50%) cartilage.
Results: Mean Dice Similarity Coefficients (DSCs) of automated vs. manual segmentation were 0.89 in the tibia and 0.83 in the femur. Deep layer femorotibial T2 was similar for automated (45.7±2.6ms) and manual (45.7±2.7ms) analyses, but superficial T2 was slightly overestimated by the automated (53.2±2.2ms) vs. manual (51.1±2.1ms) approach. Correlation of T2 by both approaches was r≥0.91 for deep and ≥0.86 for superficial layers. A significant (p<0.05) longitudinal T2 increase was seen in the deep layer of the lateral femur, but not in other regions, using both methods. No relevant differences in baseline or longitudinal T2 were detectable between the ACL-injury groups vs. the healthy volunteers, using either method.
Conclusions: An automated, CNN-based approach of cartilage T2 analysis is clinically reliable at 1.5T for analyzing cartilage composition with MRI, cross-sectionally and longitudinally.