A fully-automated technique for cartilage morphometry in knees with severe radiographic osteoarthritis - Method development and validation

Abstract

Objective: Denuded areas of subchondral bone (dAB) pose a challenge for fully automated segmentation of articular cartilage and subchondral bone in knees with severe radiographic osteoarthritis using convolutional neural networks (CNNs). Here we propose an automated post-processing relying on a selection-based multi-atlas registration for reconstructing the total area of subchondral bone (tAB) to overcome this issue. We evaluate the agreement, accuracy and longitudinal sensitivity to cartilage change of this novel methodology.

Design: CNN-based models were trained using manual cartilage segmentations from sagittal DESS and coronal FLASH MRI of knees with radiographic (KLG2-4) or severe radiographic osteoarthritis (KLG4 only). These were then applied to KLG4 test knees with manual cartilage segmentations. Automated post-processing was applied to reconstruct missing parts of the tAB and to refine the segmentations, particularly for dABs. The agreement and accuracy of automated cartilage analysis were evaluated using Dice Similarity Coefficients (DSC) and Bland-Altman analyses; sensitivity to one-year change was assessed using the standardized response mean (SRM).

Results: Stronger agreement (DSC 0.80 ​± ​0.07 to 0.89 ​± ​0.05) and lower systematic offsets for cartilage thickness (1.2 ​%-8.4 ​%) and tAB area (-0.4 ​%-4.3 ​%) were observed for CNNs trained on KLG2-4 rather than KLG4 knees; overall, results were superior to those without registration-based post-processing. Sensitivity to change was greatest for manual segmentation of DESS (SRM ​≥ ​-0.69; automated: ≥-0.56) and for automated segmentation of FLASH (≥-0.74; manual ≥-0.44).

Conclusion: CNN-based segmentation combined with registration-based post-processing for accurate delineation of tABs/dABs substantially improves fully-automated (longitudinal) analysis of cartilage and subchondral bone morphology in knees with severe radiographic osteoarthritis.

Keywords: Convolutional neural network; Fully-automated analysis; Imaging; Osteoarthritis; Severe radiographic OA.

Fig. 1

Flow chart illustrating the automated segmentation and post-processing. The selected example shows the automated segmentation of the central medial femoral condyle. The automated segmentation relies on 2D U-Nets for separately segmenting the cartilage and the total area of subchondral bone (tAB) from MRI. The post-processing uses an atlas of manually segmented tABs that is registered to the automatically segmented tAB and the best-matching reference tAB is chosen for the reconstruction of the tAB. After the reconstruction of the tAB, the tAB is combined with the cartilage segmentation, before the combined segmentation is checked for segmentation errors and automatically cleaned.

Fig. 2

Illustration of segmentations with very low and very high agreement vs. manual reference segmentations from both DESS and FLASH MRI. The numbers provide the Dice Similarity Coefficient (DSC) for the cartilages (MT/LT: medial/lateral tibia; cMF/cLF: central medial/lateral femoral condyle). The yellow arrows indicate areas with severe differences between automated and manual segmentations.

Fig. 3

Bland & Altman plots illustrating the accuracy of automated cartilage thickness measures vs. cartilage thickness measures from manual reference segmentations for the three models (KLG4_only (BL), KLG4_only (BL ​+ ​FU), KLG2-4(BL)) from both DESS and FLASH MRI. The numbers provide the mean differences between cartilage thickness measured form automated and manual segmentations and the 95 ​% levels of agreement.

Fig. 4

Bar graphs showing the systematic offset of automated cartilage thickness measures in the more severely affected medial compartment from the three models (KLG4_only (BL), KLG4_only (BL ​+ ​FU), KLG2-4(BL)) for both DESS (top) and FLASH MRI (bottom). The dotted lines indicate the systematic offsets observed with the previous automated segmentation technique using a relatively small KLG 4 model, a relatively large KLG 4 model, and a model trained on knees with all grades (KLG2-4) representing ROA [21]. Please note that the previous technique used by Wisser et al. [21] did not include the registration-based post-processing for reconstructing the total area of subchondral bone (tAB).