Work

Comparison of Fully Automated Vs Manual Image Segmentation for the Assessment of Articular Cartilage T2 Relaxation Times in Klg0 Knees With and Without Cartilage Damage - on Behalf of the Oa-bio Consortium

Year: 2023

Type: article

Abstract: MRI-detected cartilage damage is associated with elevated cartilage T2 relaxation times on MRI. Further, cartilage damage is more frequently observed in knees at higher risk of developing radiographic knee osteoarthritis (OA). Detection of such early changes by means of MRI may be a useful method to identify participants for future observational or interventional OA studies. The aim of this study was to compare a U-Net-based, fully automated cartilage segmentation pipeline vs. manual, quality-controlled cartilage segmentations for determining differences in laminar cartilage T2 in Kellgren-Lawrence Grade (KLG) 0 knees with and without MRI Osteoarthritis Knee Score (MOAKS) cartilage damage. The fully automated U-Net-based image analysis pipeline was trained using sagittal multi-echo spin-echo (MESE) MRIs from the Osteoarthritis Initiative (OAI), for which manual quality-controlled segmentations of the respective structures were available. U-Nets for medial (MFTC) and lateral (LFTC) femorotibial cartilage segmentation were trained on all seven echoes of the MESE MRIs of 92 OAI healthy reference cohort (HRC) participants. A third U-Net was trained on bone segmentations of 60 OAI HRC knees. The bone segmentation served to algorithmically determine the weight-bearing femoral region of interest and to select the MRI slices that required cartilage segmentation. MFTC and LFTC U-Nets were then applied to the selected MRI slices of 123 KLG0 knees of the OAI incidence cohort (small sample) with, and 618 KLG0 knees of the same cohort without manual cartilage segmentations available (total of 741 knees = full sample). Automated post-processing was used to identify and correct implausible segmentations. Differences in superficial and deep layer cartilage T2 were compared between knees with vs. without MOAKS cartilage damage in the small sample (both U-Net and manual segmentations) and full sample (U-Net segmentation only). Mean differences were considered statistically significant when the 95% confidence interval did not include 0. MOAKS readings were performed by a very experienced MSK radiologist. Cartilage damage was defined as MOAKS cartilage score >0 in any of the medial or lateral tibial MOAKS subregions, or in the central medial or lateral femoral MOAKS subregions. Effect sizes were expressed using Cohen's D (d). The full sample comprised knees of 415 women and 326 men (age [mean ±SD]: 60±9 years, BMI: 27±4 kg/m²). Of these, 408 had no cartilage damage, 115 only medial damage, 149 only lateral damage and 69 had medial and lateral damage. Mean cartilage T2 values for the medial and lateral femorotibial compartments, calculated from the manual and automated segmentations in the small sample, as well as from the automated segmentations in the full sample are shown in Figure 1. In the MFTC, superficial layer T2 was longer in knees with cartilage damage than in those without for manual and automated segmentations of the small sample (d [95% CI] = 0.63 [0.26, 0.99]; d=0.58 [0.22, 0.95]), and also for automated segmentations of the full sample (d = 0.70 [0.53, 0.87]). Deep layer T2 did not differ between knees with vs. without cartilage damage for the automated cartilage segmentation of the large sample and the automated and manual cartilage segmentations of the small sample (d=0.04 to 0.09). Manual cartilage segmentations of the LFTC yielded greater T2 for both the deep (d = 0.48 [0.12, 0.85]) and superficial layers (d = 0.66 [0.29, 1.03]) in knees with vs. without cartilage damage. Results for U-Net segmentations were d=0.59 [0.22, 0.95] (small sample) and d=0.43 [0.27, 0.59] (full sample) for the deep layer and d=0.66 [0.29, 1.02] (small sample) and d=0.59 [0.43, 0.75] (full sample) for the superficial layer, respectively. The fully automated U-Net-based cartilage segmentation pipeline was at least as sensitive to differences in laminar cartilage T2 between KLG 0 knees with vs. without MOAKS cartilage damage as quality-controlled, manual cartilage segmentations. Hence, this approach shows great potential for laminar cartilage T2 analyses and for application to large samples. This work was funded as part of the OA-BIO Eurostars-2 project (E! 114932). FE, WW, SM and AW are part-time employees of Chondrometrics GmbH; FE, WW, and SM are co-owners of Chondrometrics GmbH. FE has provided consulting services to Merck KGaA, Kolon-TissueGene, and Novartis. FB is co-owner of 4Moving Biotech. F.W.R. and A.G. are shareholders of BICL, LLC. A.G. is consultant to Pfizer, Kolon TissueGene, Novartis, AstraZeneca, Coval, Medipost and ICM. GND has no conflicts of interest to declare. CORRESPONDENCE ADDRESS: [email protected]

Source: Osteoarthritis Imaging

DOI: 10.1016/j.ostima.2023.100130