ABSTRACTThis work envisages developing an automated computer workflow to locate the landmarks like knee center, tibial plateau, tibial and femoral axis to measure Femur‐Tibia Angle (FTA), Medial Proximal Tibial Angle (MPTA), and Hip Knee Ankle Angle (HKAA) from the pre‐ and post‐operative x‐rays. In this work, we propose a variant of semantic segmentation model (vSegNet) for the segmentation of the knee and tibia gap for extracting important features used in the automated workflow. Since femur tibia gap is a small region as compared to the complete x‐ray image, it poses severe class imbalance issue. Using a combination of the Dice coefficient and Hausdorff distance as a compound loss function, the proposed neural network model shows better segmentation performance as compared to state‐of‐the‐art segmentation models like U‐Net, SegNet (with and without VGG16 pre‐trained weights), VGG16, MobileNetV2, Pretrained DeepLabv3+ (Resnet18 weights), and Pretrained FCN (VGG16 weights) and different loss functions. We subsequently propose computer methods for feature recognition and prediction of landmarks at femur, tibial and knee center, the side of the fibula and, subsequently, the various knee joint angles. An analysis of sensitivity of segmentation accuracy on the accuracy of predicted angles further substantiate the efficacy of the proposed methods. Dice score of U‐Net, Pretrained SegNet, SegNet, VGG16, MobileNetV2, Pretrained DeepLabv3+, Pretrained FCN, vSegNet with cross‐entropy loss function and vSegNet with compound loss function are observed as , , , , , , , and respectively. Using the proposed network vSegNet and the automated workflow, we obtained an Intraclass correlation of 0.999, 0.994, and 0.997 for the FTA, MPTA and HKAA measurements, respectively, with the ground truth. FTA, MPTA, and HKAA measurements using the proposed automatic pipeline positively correlated with the expert's measurement. The proposed vSegNet with compound loss function handles the challenges posed by class imbalance and obtains the best results as compared to other networks and loss functions tested in the work and also in comparison with contemporary works described in literature.
Read full abstract