nnU-NET BASED 3D SEGMENTATION OF HIGH-RESOLUTION KNEE CT SCANS FOR THE ANALYSIS OF SHAPE AND INTRA-ARTICULAR...

High-resolution knee CT scans provide fine detail about the 3D shape of the bones in the knee, but manually segmenting multiple slices in fine detail can be very tedious. Intra-articular mineralization quantification could potentially be a useful imaging biomarker, but its quantification would require masking out the signal from bone mineral. Therefore, an automated method of accurately defining the extent of the femur, tibia, and patella on knee CT potentially has multiple uses.

To develop an automated tool to segment the femur, tibia, and patella from knee CT scans.

From 4,014 knee CT scans acquired in the MOST Study at Year 12, 234 knees were selected to represent a range of osteoarthritis (OA) severity based on tibiofemoral (TF) and patellofemoral (PF) Kellgren-Lawrence grades (KLG). Knees were categorized as" (1) no OA (TF KLG < 2, PF KLG < 2) (2) TF OA only (TF KLG ≥ 2, PF KLG < 2) (3) PF OA only (TF KLG < 2, PF KLG ≥ 2) (4) mixed OA (TF KLG ≥ 2, PF KLG ≥2 2) TF OA cases were further classified as medial or lateral using OARSI joint space narrowing grades. The sample was balanced across OA categories, TF compartment involvement, and KLG distributions. Femur, tibia, and patella were manually segmented and used to train and validate a deep learning model based on the nnU-Net framework. Images were acquired with a standardized protocol and left unnormalized, but trimmed to a standardized 450 axial slices centered around the femoral epicondyles to give 3D image volumes of size 512 × 512 × 450. Images were split into training and validation sets for 5-fold cross-validation, but stratified to ensure equal numbers of knees from each OA class in each of the 5 splits. We also made sure that if a study participant contributed 2 knees to the sample, both knees would be in the same split. For the 5-fold cross-validation, we used a 3d nnU-Net v2 cascaded architecture combining coarse localization trained on an entire low-resolution 3D volume with a subsequent patch-based full-resolution refinement.

Initial runs of the low-resolution 3D training showed no signs of overfitting or underfitting by 500 epochs, so the entire cascade was performed using each of the 5 training sets, and then the 5 models were used to predict segmentations for their matching validation sets. Table 1 shows the results for agreement between the manual segmentations and the predicted segmentations presented as DICE coefficients. Figure 1 shows an example of a predicted segmentation.

The nnU-NET framework provides a simple and reliable method for developing a deep learning model for segmenting femur, tibia and patella from CT scans of the knee. This model was developed on high resolution unilateral reconstructions. By mapping these segmentations will be mapped back onto lower resolution bilateral reconstructions a matching model which will be able to segment both left and right knees from a single reconstruction, albeit at lower resolution can also be developed.