A FULLY-AUTOMATED TECHNIQUE FOR KNEE CARTILAGE AND DENUDED BONE AREA MORPHOMETRY IN SEVERE RADIOGRAPHIC KNEE OA – METHOD DEVELOPMENT AND VALIDATION

Automated cartilage segmentation using convolutional neural networks (CNN) has been shown to provide moderate to high accuracy in comparison with gold-standard manual approaches. It also displays similar sensitivity to longitudinal change and to between-group differences in change as has been reported for manual analysis [1-3]. Denuded areas of subchondral bone (dAB) provide challenges and impair the accuracy of automated cartilage segmentation in knees with severe radiographic OA (KLG 4). The reason is that CNNs are trained to detect cartilage, but encounter “difficulties” to properly segment areas where cartilage is lost entirely. CNNs therefore often segment cartilage cover in some areas of actual full thickness loss or ignore dABs entirely. This was observed to result in an overestimation of cartilage thickness and an underestimation of dABs in knees with severe OA [4]. To improve CNN-based automated segmentation in severely osteoarthritic knee cartilage by using an automated post-processing algorithm that relies on a multi-atlas registration for reconstructing the total area of subchondral bone (tAB). We evaluate the agreement, accuracy and longitudinal sensitivity to cartilage change of this new methodology. Sagittal DESS and coronal FLASH MRIs were acquired by the Osteoarthritis Initiative (OAI). 2D U-Net models were trained for both MRI protocols using manual cartilage segmentations of knees with radiographic OA (KLG2-4, n training / validation set: 86/18 knees, baseline scans only) or severe radiographic OA (KLG4, n training/ validation set: 29/6 knees. These were trained either from baseline scans only [KLG4 BL ] or from baseline and follow-up scans [KLG4 BL+FU ]. The trained models were then applied to the test set comprising 10 KLG4 knees with manual cartilage segmentations from both DESS and FLASH MRI available and to n=125/14 knees with manual cartilage segmentations from either DESS or FLASH MRI available. Automated, registration-based post-processing was applied to reconstruct missing parts of the tAB and to refine the segmentations (Fig. 1), particularly in areas of denuded bone. The agreement and accuracy of automated cartilage analysis were evaluated in the test set for individual cartilages using Dice Similarity coefficients (DSC), correlation analysis, and by determining systematic offsets between manual and automated analysis. The sensitivity to one-year change was assessed using the standardized response mean (SRM) across the entire femorotibial joint in 104/24 (DESS/FLASH) knees with manual baseline and follow-up segmentations. The strongest agreement (DSC 0.80±0.07 to 0.89±0.05) and lowest systematic offsets for cartilage thickness (1.2% to 8.5%) were observed for CNNs trained on KLG2-4 rather than KLG4 knees. Similar observations were made for dABs (-40.6% to 3.5%) and total subchondral bone area (-0.4% to 4.3%). Fig. 2 displays the offsets for cartilage thickness together with the offsets previously observed without the registration-based post-processing. Cartilage thickness obtained from the KLG2-4 model was strongly correlated with that from manual segmentations (r=0.82 to r=0.97) whereas a moderate to strong correlation was observed for dABs (r=0.52 to r=0.92). The sensitivity to change across the entire femorotibial joint was greatest for manual segmentation of DESS (SRM -0.69; vs. automated: -0.28 to -0.56) but on the other hand for the automated segmentation of FLASH (-0.47 to -0.67; vs. manual = -0.44, Fig. 3) MRI. CNN-based segmentation combined with registration-based post-processing for accurate delineation of tABs/dABs substantially improves fully-automated analysis of cartilage and subchondral bone morphology in knees with severe radiographic OA when compared to a previously used fully-automated approach without such post-processing [4]. This finding is consistent for two different MRI contrasts (DESS and FLASH) and orientations (sagittal and coronal). Our results additionally show that the more general (KLG2-4) model is well suited for automated segmentation of KLG4 knees, eliminating the need for a KLG4-specific model.