Articular cartilage transverse relaxation time (T2) on MRI has been observed to reflect collagen integrity, orientation, and hydration, and to be associated with cartilage histological grading, mechanical properties, and early knee OA status. Previously, we reported high agreement of cartilage segmentations obtained from multi echo spin echo (MESE) MRI, using fully automated deep learning methods, in comparison with manual ones. As these results were based on 3T MRI (of the OAI), we here explored the agreement between fully automated segmentation of MESE cartilage at 1.5T. To analyze the agreement between automated, U-Net-based segmentation of femorotibial cartilage from MESE by convolutional neural networks vs. manual expert segmentation with quality control of the same images, and to compare laminar femorotibial cartilage T2 between both approaches. We studied 20 ACL-deficient patients with persistent dynamic knee instability (non-copers), 22 without instability (copers), 13 with surgical ACL reconstruction, and 16 healthy controls. Further patient characterizations are described in a parallel abstract at this conference. Sagittal MESE MRIs were acquired at 1.5T (Siemens Avanto; slice thickness/spacing 3/3.5mm, in-plane resolution 0.31mm, TR 1500ms, TE 9.7/19.4/29.1/38.8/48.5/58.2/67.9ms), at baseline (n=71) and 1 year later (n=55). Manual cartilage segmentation was done by experienced readers, with quality control by an expert (Fig. 1). The U-Net was trained using medial and lateral MRIs from the same scanner (training/ validation set n=50/9) obtained in volleyball athletes of different age groups, and in patients with posterior cruciate ligament (PCL) surgery, segmented by the above readers. Training of the U-net was performed for both all 7 echos and only the 1st echo. Automated U-Net segmentation was then applied to the current study MRIs, without manual intervention or correction. Yet, automated post-processing was employed to correct obvious segmentation errors. The agreement between manual and automated U-Net-based segmentations was evaluated using the Dice Similarly Coefficient (DSC), and by evaluating systematic differences and correlations in cartilage T2 of the 50% superficial and 50% deep layer. When all echoes were used for training, the overall DSC was 0.89 for MT/LT, and 0.83 for cMF/cLF. When only the 1st echo was used,DSCs were 0.87/0.88 and 0.81/0.79. Automated analysis overestimated the segmented tissue volume significantly in most regions, with correlations ranging from 0.93-0.96 for all echoes, and from 0.87-0.94 for the 1st echo. Deep layer T2 across the joint was 45.7ms for manual analysis, 45.7ms with all echoes, and 46.1ms with the 1st echo model; superficial layer T2 was 52.1, 53.2 and 54.4ms. There were statistically significant (albeit small) differences of automated vs. manual analysis across most regions for the superficial layer, and for LT and cLF in the deep layer. The correlation of deep layer T2 across the plates ranged from 0.91-0.99 for all echoes model, and from 0.85-0.98 for the 1st echo model, and that of the superficial T2 from 0.86-0.97, and from 0.73-0.81. Fully automated (U-Net-based) laminar analysis of femorotibial cartilage T2 appears feasible at 1.5T, albeit the agreement for T2 at 3T, and that for cartilage thickness was previously reported to be higher. The agreement with manual analysis was greater when training a model with 7 echoes than with the 1st echo only, and it was greater for the deep than for the superficial layer. No significant change in T2 was observed over 1 year; thus, longer intervals may be required for longitudinal validation.
Read full abstract