AUTOMATED KNEE CARTILAGE SEGMENTATION FOR T2 QUANTIFICATION USING HIGH-RESOLUTION MRI: VALIDATION IN THE NACOX ACL...

T2 mapping enables early detection and monitoring of cartilage degeneration following ACL injury, yet reliable quantification requires accurate segmentation of cartilage from high-resolution MRI data. Manual segmentation of cartilage is time-consuming and evaluator-dependent, limiting its use in large cohort studies. Deep learning provides a scalable alternative, but validation using high-resolution MRI data from post-traumatic OA cohorts is limited.

1) To validate automated deep learning-based cartilage segmentation as a substitute for manual segmentation in T2 quantification using high-resolution clinical MRI data from the NACOX ACL injury cohort. 2) To compare segmentation performance and failure patterns between the second echo image of the multi-echo train and the T2 parametric map training inputs.

From the NACOX cohort, a multicenter study of patients aged 15-40 years with acute ACL injury, we used baseline data from 129 patients enrolled at Linköping University Hospital, Sweden. After excluding patients with missing MRI scans, image artifacts, and missing annotations, 117 patients (231 knees) were included. MRI was performed on a Philips Ingenia 3T scanner using a multi-echo spin-echo sequence (TE=10-80ms, 8 echoes; in-plane resolution 0.29 × 0.29mm; slice thickness 3mm; 26 sagittal slices per volume). T2 maps were derived via monoexponential fitting, including values in the range of 10-90ms. Manual segmentation of femoral and tibial cartilage (FC, TC) was performed on 6 slices per scan, centered within the medial and lateral compartments (3 slices each). Segmentations were performed by a trained observer supervised by an expert with 8 years of experience (mean intrareader RMS-CV: 2.0% FC, 2.4% TC). The dataset was split into 93 patients (183 scans) for training and 24 patients (46 scans) as a hold-out test set. We trained an ensemble of Swin-UNETR models (Hatamizadeh et al. Lect Notes Comput Sci 12962 2022) using 5-fold cross-validation. Two input data settings were compared: the second echo image of the echo train ("Echo 2", TE=20ms) and T2 parametric maps (Ibrahim et al. Osteoarthritis Cartilage 33(S1) 2025).We assessed segmentation accuracy using Dice score, average symmetric surface distance (ASSD), and 95th-percentile Hausdorff distance (HD95), and evaluated downstream validity via Bland-Altman agreement analysis of average T2 values between automated and manual segmentations tissue-wise. Unless otherwise stated, values are reported as mean ± SD.

Both input datasets achieved high segmentation accuracy on the test set (Table 1). "Echo 2" and T2 map showed similar Dice scores (FC: 0.887±0.022 vs 0.894±0.024; TC: 0.899±0.019 vs 0.903±0.018) and were comparable with T2 agreement (FC: 0.22ms [95% LoA: -0.47, 0.92ms] vs 0.07ms [95% LoA: -0.60, 0.74ms]; TC: -0.09ms [95% LoA: -1.08, 0.90ms] vs 0.14ms [95% LoA: -0.70, 0.97ms]). Boundary errors remained within 1-2 voxels for both datasets (HD95 ≤0.42mm), with "Echo 2" segmentation showing marginally better boundary accuracy for FC (HD95: 0.371±0.216 vs 0.415±0.520mm) and comparable accuracy for TC (HD95: 0.322±0.091 vs 0.346±0.109mm). Slice-level Dice distributions were right-skewed with medians above 0.890 for femoral and tibial cartilage; a small fraction of outlier slices fell below 0.800, predominantly at locations with ambiguous bone-cartilage boundaries or reduced tissue contrast (Figure 1). The high spatial resolution of this dataset (>10 voxels across cartilage thickness) enabled reliable qualitative error characterization beyond average metrics. Failures with "Echo 2" input data were characterized by under-segmentation at the posterior femoral boundary, consistent with reduced bone-cartilage contrast at that region. T2 map failures showed a distinct pattern of over-segmentation along the femoral surface, possibly reflecting model sensitivity to high-contrast periarticular tissue in noisy parametric maps. Tibial errors were generally minor but more pronounced with Echo 2 input data, while T2 map input data showed better segmentation for the same cases.

Automated knee cartilage segmentation for T2 quantification demonstrated reliable performance in a novel ACL injury cohort with high-resolution MRIs, with negligible bias and comparable limits of T2 agreement for both input datasets. The two training inputs showed practically similar overall performance; minor differences in femoral boundary accuracy and T2 bias were observed, yet they were negligible when confidence intervals were considered. These findings, based on the highest-resolution T2 mapping protocol reported to date for this application, establish a robust reference for future studies.