Key-point estimation of knee X-ray images using a parallel fusion decoding network

High tibial osteotomy (HTO) is a knee preservation procedure used to treat osteoarthritis of the knee. Identifying the hinge point, surgical point, and Fujisawa point in the patient's knee X-ray before surgery is a critical task. The aim of this study was to propose an artificial intelligence-based method to effectively help surgeons select the location of these landmark points, which provides important reference for subsequent surgery.

We proposed PFDNet (parallel fusion decoding network), a novel convolutional neural network for key-point estimation of knee X-rays. PFDNet employs Res2Net for feature extraction in the network encoding phase and two partial decoders connected in parallel in the network decoding phase to finely aggregate the multiscale feature information produced by Res2Net. A total of 1842 knee X-ray images were trained, validated and predicted by PFDNet to determine whether the network could accurately detect key-points in the HTO surgical plan.

At the hinge point, surgical point, and Fujisawa point, the average error and standard deviation from the calibration value in the PFDNet test results were 2.06 ± 1.165 mm, 2.713 ± 1.457 mm, and 2.015 ± 1.304 mm, respectively. This method exhibits superior performance compared with four convolutional neural network models that are also based on encoding and decoding frameworks: U-Net, ResUnet, SegNet, and FCN.

The hinge point, surgical point, and Fujisawa point can be clearly selected by PFDNet from knee X-ray images and is locked to the millimeter level. The results show that the proposed artificial intelligence-based strategy can be instrumental in preoperative HTO planning.