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

Area of Science:

• Orthopedic surgery
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• High tibial osteotomy (HTO) is a knee preservation technique for osteoarthritis.
• Accurate identification of surgical landmarks (hinge point, surgical point, Fujisawa point) on knee X-rays is crucial for HTO.
• Current methods for landmark identification can be challenging and time-consuming.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI)-based method for precise landmark identification in preoperative HTO planning.
• To assist surgeons in accurately selecting key points on knee X-rays.

Main Methods:

• A novel convolutional neural network, PFDNet (parallel fusion decoding network), was developed for key-point estimation on knee X-rays.
• PFDNet utilizes Res2Net for feature extraction and parallel decoders for multiscale feature aggregation.
• The model was trained and validated on 1842 knee X-ray images.

Main Results:

• PFDNet achieved average errors of 2.06 ± 1.165 mm (hinge point), 2.713 ± 1.457 mm (surgical point), and 2.015 ± 1.304 mm (Fujisawa point).
• The AI method demonstrated superior performance compared to U-Net, ResUnet, SegNet, and FCN models.
• Landmark identification was accurate to the millimeter level.

Conclusions:

• PFDNet effectively identifies critical landmarks for HTO surgery from knee X-ray images.
• The proposed AI-based strategy significantly aids in preoperative planning for high tibial osteotomy.
• This technology offers a precise and reliable tool for orthopedic surgeons.