AUTOMATIC FEMOROTIBIAL JOINT SPACE PATCH SELECTION FROM WEIGHT-BEARING CONE-BEAM COMPUTED TOMOGRAPHY BONE...

Weight-bearing cone-beam CT (CBCT) enables evaluation of the knee joint under physiological load in a standing position. It has been used to assess OA-induced structural changes in the joint with a higher sensitivity compared to conventional radiography. However, labor-intensive manual bone and femorotibial joint space patch (FT-JSP) segmentation is currently required to quantify the 3D anatomy and joint space width (JSW) of the load-bearing region with CBCT. Here, we implemented a deep learning model for automatic and accurate FT-JSP selection.

To develop a deep learning model to segment FT-JSP from 2D map representations derived from bone segmentations.

The dataset comprised weight-bearing CBCT scans from a weight-loss surgery study acquired at multiple time points. Imaging was performed using the Planmed Verity system with 0.2 mm isotropic voxel size, 96 kVp tube voltage, 12 mA current, and 6 s acquisition time. Of 264 scans from 86 patients, four were excluded due to errors in the generated femur/tibia objects, leaving 260 scans for analysis. The femur, tibia, and FT-JSP were manually segmented using Stradview, and the FT-JSP segmentation was used as a gold standard. From these segmentations, 2D geometric representations were derived, including a heightmap calculated 10 mm above the highest position of the femur and an FT distance map. Data were split at the patient level into training (n=176), validation (n=31), and hold-out test (n=53) sets to prevent leakage across repeated time points. A residual 2D U-Net implemented in the PyTorch-based MONAI framework was trained to segment the FT-JSP from these derived 2D maps, and the prediction was thresholded at 0.8 to obtain a binary segmentation mask. The performance of the method was evaluated using the Dice score, Jaccard index, precision, and recall. Finally, the segmentation was projected on the femur 3D surface to evaluate the 95^th percentile Hausdorff distance (HD95).

The automatic FT-JSP segmentation achieved a Dice score of 0.90 ± 0.03, 0.89 ± 0.05, 0.88 ± 0.03, and a Jaccard index of 0.83 ± 0.04, 0.80 ± 0.07, 0.79 ± 0.05 for training, validation, and hold-out sets, respectively (Table 1). On the hold-out test set, precision and recall were 0.89 and 0.89, respectively. After reconstructing the automatically obtained FT-JSP masks into a 3D surface representation, the HD95 to the manual segmentation was 1.96 ± 0.21 mm, with a representative qualitative example shown in Figure 1.

A residual 2D U-Net can automatically segment the FT-JSP from weight-bearing CBCT-derived geometric representations. These results support the feasibility of the method for automated FT-JSP identification and show that the predicted 2D regions can be reconstructed into 3D space with good anatomical agreement. The next steps in development include validating an end-to-end automated pipeline for FT joint quantification. This approach would also benefit from test-retest evaluation to assess the stability of patch localization across repeated scans and differences in knee flexion angles.