Development of an Automated Mass-Customization Pipeline for Knee Replacement Surgery Using Biplanar X-Rays

Abstract

For standard ‘off-the-shelf’ knee replacement procedures surgeons use X-ray images to aid implant selection from a limited number of models and sizes. This can lead to complications and the need for implant revision due to poor implant fit. Customised solutions have been shown to improve results but require increased preoperative assessment (Computed Tomography or Magnetic Resonance Imaging), longer lead times and higher costs which have prevented widespread adoption. To attain the benefits of custom implants, whilst avoiding the limitations of currently available solutions, a fully automated mass-customisation pipeline, capable of developing customised implant designs for fabrication via additive manufacturing from calibrated X-rays, is proposed. The pipeline uses convolutional neural networks to extract information from bi-planar X-ray images, point depth and statistical shape models to reconstruct the anatomy, and application programming interface scripts to generate various customised implant designs. The pipeline was trained using data from the Korea Institute of Science and Technology Information. Thirty subjects were used to test the accuracy of the anatomical reconstruction, ten from this dataset and a further twenty independent subjects obtained from the Osteoarthritis Initiative. An average root mean squared error of 1.00 mm was found for the femur test cases and 1.07 mm for the tibia. 3D distance maps of the output components demonstrated these results corresponded to well-fitting components, verifying automatic customisation of knee replacement implants is feasible from 2D medical imaging.

Applying machine learning methods to enable automatic customisation of knee replacement implants from CT data

“…Furthermore, the SSM base shapes were ‘idealised’ and smoothed so outputted models were suitable for use in designing custom implants. This approach is consistent with Burge et al 12 , 13 .…”

“…Finally, although generic TKR components were used in this study, it would be possible to adapt the pipeline to work with commercially available designs. Applying the CT-based pipeline framework to other medical implants, such as uni-condylar knee replacements 12 , hip replacements 32 , or for predicting the best non-customised TKR implant sizes 33 , could also be explored.…”

“…Moreover, authors have outlined how the design of custom medical devices, including cranial implants 6 , 10 and ocular prostheses 11 , can be achieved automatically from CT data via similar techniques. Burge et al used various machine learning techniques to develop a ‘fully automated mass-customisation pipeline’ for producing TKR components from bi-planar X-ray images 12 . The authors however found that, despite multiple benefits of using X-rays as input (reduced radiation exposure, cost, imaging time etc.…”

“…This study therefore outlines a new approach for automating CT-based custom TKR implant design. Machine learning-based medical image processing methods are exploited to enable a ‘mass-customisation’ pipeline—the automated design of customised products for the masses 12 . The ambition is that such a solution would enable customised implants to be fabricated faster and with lower cost, making the numerous potential benefits of customisation 16 , 17 more obtainable.…”

A computational tool for automatic selection of total knee replacement implant size using X-ray images

“…Like in Zheng et al (2018) and Massé and Ghate (2021) , the size prediction tool initially utilizes inputted AP and lateral X-ray images to generate 3D estimations of patients’ femur and tibia bones before component size predictions can be made. To achieve this automatically, the 2D—3D pipeline developed in Burge et al (2022a) , Burge et al (2022b) was utilized. The key aspects of the workflow, built using Python 3, are summarized below and illustrated in Figure 1 .…”