AUTOMATED CARTILAGE T2 ANALYSIS BY REGISTRATION OR BY SEGMENTATION USING CONVOLUTIONAL NEURAL NETWORKS (CNNs) – WHICH ONE MAKES THE RACE?

Abstract

INTRODUCTION

Manual cartilage segmentation from MRI is a labor-intensive process. This is particularly cumbersome in studies in which cartilage morphology is to be determined from manual segmentation of fat-suppressed, high-resolution gradient echo (GrE) sequences, and then T2 from another manual segmentation of a multi echo spin echo (MESE) sequence. To this end, we have developed a registration algorithm that uses segmentations of the cartilage from the GrE sequences, and rigidly registers these to an optimal position for extracting cartilage T2 signal from the MESE [1,2]. However, we have recently started to develope fully automated analysis technology for T2 directly from the MESE using convolutional neural network (CNN) architectures and deep learning (DL) [3].

OBJECTIVE

To compare i) T2 determined from MESE by registration of manually segmented cartilage masks from GrE and ii) T2 determined from MESE directly by fully automated segmentation using CNNs [3] vs. manually segmentation for T2 analysis in the same knees.

METHODS

We studied 39 ACL patients and 15 healthy controls, enrolled at Charité (n=54; Berlin, Germany). Sagittal 3D VIBEwe MRIs were acquired for cartilage morphometry, and sagittal 2D multi-echo spin-echo (MESE) MRIs for cartilage T2 analysis using a 1.5T Siemens Avanto MRI, at baseline and (n=53) at 1 year follow-up. Segmentation of the femorotibial cartilages was performed manually by expert readers from the 3D VIBE and 2D MESE. A multimodal approach was used to register cartilage segmentations from the VIBE to the MESE [1,2]. Automated cartilage segmentation of the MESE relied on a 2D U-Net [3] that was trained on all 7 echoes from athletes and PCL patients (training/validation set n=50/9), the images being acquired on the same scanner and segmented by the same readers. Agreement between registered and automated vs. manual cartilage segmentation was assessed using dice similarity coefficients (DSCs). Superficial and deep femorotibial cartilage T2 (each 50% thickness) were extracted from the segmentations. Baseline cartilage T2 and 1-year change were compared between methods, using Pearson correlation coefficients, mean differences, and 95% CIs.

RESULTS

In the deep cartilage layer, baseline T2 derived from automated (CNN) segmentation was very similar to that of the manual expert segmentation on the same images, with mean differences of 0.1ms, and correlations of r=0.97-98 across compartments (Table 1). Deep T2 values obtained from registration were longer than those from manual segmentations, with correlations of 0.12-0.13. Superficial T2 (Table 1) was approx. 6-7ms longer than that in the deep layer across all methods. The CNN method overestimated T2 by about 1.2ms (r=0.91-93), and the registration method by about 8ms (r<0.13). The longitudinal results confirmed superiority of the direct CNN segmentation (Table 2).

CONCLUSION

Direct automated segmentation of MESE using CNN-based segmentation [3] yields highly accurate results of T2, a measure of cartilage composition, in deep and superficial cartilage laminae. This here applied cross-sectionally and longitudinally at 1.5T, with similar results obtained at 3T [4]. The alternative of extracting T2 via registration [1] of cartilage morphology masks from GrE displayed less accurate results. Although the registration algorithm was previously shown relatively accurate [1], and predictive of progression in the OAI FNIH sample [2], these results were obtained at 3T. Further, a “peel factor” was introduced to limit GrE mask in the depth in order to fit the MESE segmentations. A uniform peel factor of 30% was used, although variation was noted between cartilage locations and participants. The same factor (30%) was used here, without optimization to the VIBEwe sequence and 1.5T. Yet, in future studies we recommend direct extraction of T2 using CNNs rather than trying to further optimize registration-based techniques as a proxy of cartilage composition.