GENTLE(WO)MAN'S DUEL!: SENSITIVITY TO CARTILAGE THICKNESS CHANGE OF CORONAL FLASH VS SAGITTAL DESS WITH FULLY AUTOMATED SEGMENTATION

Abstract

INTRODUCTION

Observational studies applying MRI-based quantitative cartilage morphometry, as well as validation studies testing fully automated cartilage segmentation, often rely on the (sagittal) double echo steady state (DESS) sequence from the OAI. However, almost all current multicenter trials evaluating putative disease-modifying OA drugs (DMOADs) use conventional spoiled gradient echo MRI (e.g. FLASH), given its broader availability across worldwide vendor and MRI scanner platforms. A comparison of cartilage loss between coronal FLASH and sagittal DESS using manual segmentation was evaluated in a small sample (n=80) [1], but it is unclear which of both protocols is more sensitive in detecting differences between knees with and without structural progression, particularly in consideration of potential bias from subjective reader preferences.

OBJECTIVE

i) To directly compare the sensitivity to longitudinal change of cartilage morphometry between both MRI protocols in the FNIH-1 OA Biomarkers Consortium [2,3]; and ii) to compare the discrimination of change between progressor and non-progressor knees, both using a convolutional neural network (CNNs) deep learning (DL) algorithm [3,4] for fully automated cartilage segmentation.

METHODS

Coronal FLASH and sagittal DESS CNNs were trained (2D U-Net) using 86 OAI knees with radiographic OA that had manual cartilage segmentations from both MRI sequences [4]. Both models displayed high agreement (Dice Similarity Coefficient) and good accuracy of cartilage thickness metrics in a ROA validation/test set (n=18/18), compared with manual segmentation [2]. FLASH MRI had been acquired in one of both knees. Therefore, the current analysis focused on 309 (304 right and 5 left) knees from the FNIH-1 sample [3]: the CNN models were applied to baseline & 2-year follow-up MRIs [3] of: 100 combined progressor knees (both radiographic [>0.7mm JSW loss] & pain progression between baseline and year >2), 104 non-progressor knees, 53 knees with isolated radiographic, and 52 with isolated pain progression. Medial femorotibial (MFTC) cartilage thickness change was compared i) between all knees with (n=153) vs. without (n=156) radiographic progression and ii) between knees with vs. without combined progression (original OAI FNIH-1 analytic design [2]). The standardized response mean (SRM) was used as a measure of sensitivity to change, and Cohen's D as a measure of effect size for discriminating longitudinal change between both groups.

RESULTS

The MFTC cartilage thickness change using CNN segmentation in all knees with radiographic progression was –211µm (SRM=-0.78) for coronal FLASH and –133µm (SRM=-0.76) for sagittal DESS; it was –37µm (SRM=-0.25) and –13µm (SRM=-0.11) in knees without radiographic progression respectively (Fig. 1). Cohen's D for progressors vs. non-progressors was 0.80 for coronal FLASH and 0.81 for sagittal DESS (CNN), whereas it was 0.84 for manual cartilage segmentation. Comparing combined progressors vs. the rest of the cohort, Cohen's D was 0.56, 0.55, and 0.56 (Fig.1).

CONCLUSION

Coronal FLASH and sagittal DESS are equally sensitive to cartilage loss between knees with vs. without either radiographic or combined (symptomatic & radiographic) progression. Having used automated analysis, these results are independent of potential reader preference and bias. The magnitude of change was greater with FLASH than DESS, but did not translate into greater sensitivity or better discrimination, because of the greater SD and the greater magnitude of change with FLASH in non-progressor cohorts. A limitation of the study is the smaller sample size (n=309) relative to the full FNIH1 sample, because FLASH was only acquired in one of both knees. Further, the results of FLASH cannot be compared with ground truth, because manual segmentation was only available for DESS. However, results for automated DESS segmentation were largely consistent with the manual reference standard. Hence both MRI sequences can be recommended for use in clinical trials, whichever is more straightforward to implement across different imaging sites and MRI vendors.