VALIDATION OF A FULLY AUTOMATED CARTILAGE SPIN-SPIN (T2) RELAXATION TIME ANALYSIS WORKFLOW FROM QUANTITATIVE DESS (QDESS) MRI

Abstract

INTRODUCTION

Cartilage T2 is commonly measured by multi-echo spin-echo (MESE) MRI. MESE, however, requires long acquisition times to obtain sufficient in-plane resolution for laminar T2 analysis and does not fully cover the deep cartilage lamina [1]. Quantitative DESS (qDESS) retains both acquired echoes so that both cartilage morphology and cartilage T2 can be extracted simultaneously from a single acquisition with relatively short acquisition time [2, 3]. The qDESS thus reduces patient burden and analysis time. The MechSens trial [4] investigated the impact of unilateral anterior cruciate ligament (ACL) injury on femorotibial (FTJ) cartilage 2–10 years after injury and is the first clinical study to use qDESS MRI. Based on manual segmentations, deep layer FTJ T2 was longer in ACL than in contra-lateral (CL) non-ACL and in healthy control knees, whereas no differences in superficial layer T2 or cartilage thickness were observed.

OBJECTIVE

To technically validate an image analysis technique based on convolutional neural networks (CNN) for automated laminar cartilage T2 analysis for qDESS vs. manual segmentations, and to test whether between-knee and -group differences in deep cartilage T2 can be replicated in ACL-injured vs. control knees.

METHODS

Of 85 participants from two age groups (20–30y & 40–60y) 37 had a unilateral ACL-injury (2–10y prior to baseline: ACL_20-30: n=23, ACL_40-60, n=14). 48 healthy controls had no history of knee injury (HEA_20-30, n=24, HEA_40-60, n=24). Coronal qDESS MRIs were acquired using a 3T Siemens Prisma in both knees (resolution: 0.31mm x 0.31mm x 1.5mm, repetition time: 17ms, echo times: 4.85/12.15ms, flip angle: 15°). Manual segmentation of weight-bearing FTJ cartilages was performed with expert quality control. Automated cartilage segmentation was based on a 2D U-Net image analysis workflow. Two U-Nets were trained on both knees of odd- or even-numbered participants and were then employed to segment the knees from the other participants (even- or odd-numbered), respectively. T2 was computed for the FTJ cartilages as previously described [2]. Deep and superficial layer T2 were computed based on the position of the voxels relative to the subchondral bone and cartilage surface and were averaged across the FTJ. The segmentation agreement was evaluated using the Dice similarity coefficient (DSC). T2 was compared between segmentations using Bland & Altman plots and correlation analysis. FTJ T2 of the ACL knees was compared to T2 of uninjured CL and healthy control knees using Conover-Iman and Dunn post-hoc tests, respectively. Paired (between-knee) or unpaired (between-group) Cohen's D was used as measure of effect size of T2 differences.

RESULTS

The agreement of automated vs. manual cartilage segmentation across the four FTJ cartilages was high, with DSCs between 0.90±0.05 (central medial femur) and 0.93±0.02 (lateral tibia). Both deep and superficial layer T2 correlated strongly between techniques (r≥0.90, Table 1). Bland Altman plots indicated that the automated segmentation tended to underestimate deep T2 and to overestimate superficial T2 (Fig. 1, Table 1). In the ACL_20-30 group, deep FTJ T2 was longer in ACL than in uninjured CL knees for both manual (D=-1.24) and automated (D=-1.30) segmentations. In the ACL_40-60 group, deep FTJ T2 was longer in ACL-injured vs. CL knees for automated (D=-1.06) but not for manual segmentation (D=-0.67, Fig. 2). Comparing ACL-injured knees with those from healthy controls, deep FTJ T2 was longer in ACL_20-30 knees than in the left (but not the right) knees of the HEA_20-30 group for both manual (D left/right=-1.22/-1.15) and automated segmentations (D left/right=-1.05/-1.01, Fig. 2). Deep FTJ T2, in contrast, was longer for ACL_40-60 than for left and right HEA_40-60 knees with manual (D left/right=-1.00/-1.04), but not with automated segmentations (D left/right=-0.94/-0.99, Fig. 2). Superficial FTJ T2 did not differ between ACL-injured vs. CL knees or healthy knees using either segmentation method (Fig. 2). Results for comparisons across age groups are shown in Fig. 2.

CONCLUSION

CNN-based fully automated cartilage T2 analysis provided a very high agreement with T2 derived from manual segmentations. Importantly, it was also sensitive to ACL-injury-related prolongation in deep cartilage T2. A deep-learning-based image analysis workflow trained from high quality segmentations may thus allow to replace manual segmentations in future studies relying on qDESS MRI for cartilage T2 analyses.