Semi-quantitative scoring systems such as KIMRISS enhance granularity in knee osteoarthritis (OA) assessment and reduce inter-reader variability. However, they remain constrained by fixed, static templates that do not adapt to patient-specific anatomy. Their reliance on manual expert input also makes them time-consuming and limits scalability for large imaging datasets. These limitations underscore the need for scoring frameworks that are both anatomically adaptive and automated. To address this, we leverage data-driven modeling of anatomical variability from multi-subject MRI data to enable automated, anatomy-adaptive scoring. Our approach generates dynamic, anatomy-conforming scoring templates in sagittal knee MRI by integrating segmentation-derived bone masks with geometric analysis, including ellipse fitting for femoral head localization, PCA-based condyle axis estimation, and bounding-box partitioning for the tibia and patella. Using 80 fluid-sensitive MRI scans from the OAI and OKINADA cohorts (40 patients), and reference input from 11 expert readers, performance was evaluated against human-expert bone marrow lesion (BML) masks. Dynamic templates achieved coverage values comparable to the mean of 11 readers for the femur (0.91), tibia (0.95), and patella (0.91). Qualitative analysis confirmed consistent and bone-conforming placement. These results demonstrate that dynamic templates can provide automated, scalable, and anatomically adaptive scoring, with potential to reduce reader variability and enhance OA assessment.
Abbasi et al. (Sun,) studied this question.