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

Abstract

The aim of this study was to develop an automated pipeline capable of designing custom total knee replacement implants from CT scans. The developed pipeline firstly utilised a series of machine learning methods including classification, object detection, and image segmentation models, to extract geometrical information from inputted DICOM files. Statistical shape models then used the information to create femur and tibia 3D surface model predictions which were ultimately used by computer aided design scripts to generate customised implant designs. The developed pipeline was trained and tested using CT scan images, along with segmented 3D models, obtained for 98 Korean Asian subjects. The performance of the pipeline was tested computationally by virtually fitting outputted implant designs with 'ground truth' 3D models for each test subject's bones. This demonstrated the pipeline was capable of repeatably producing highly accurate designs, and its performance was not impacted by subject sex, height, age, or knee side. In conclusion, a robust, accurate and automatic, CT-based total knee replacement customisation pipeline was shown to be feasible and could afford significant time and cost advantages over conventional methods. The pipeline framework could also be adapted to enable customisation of other medical implants.

Figure 1

Workflow of CT-3D surface model prediction process. (a) Load DICOM, (b) classify slices, (c) identify ROI, (d) crop slices to ROI, (e) segment femur and tibia bones, (f) create contour stacks, (g) align contour stacks with SSM base shapes, (h) morph SSMs to fit contour points and create 3D model predictions.

Figure 2

Tibia and femur segmented areas by slice. Plot shows segmented bone areas plotted for each slice within the ‘Knee’ region of a CT stack. Dashed line indicates the identified tibia–femur transition point.

Figure 3

Automatic custom femur component (top row) and tibia plate (bottom row) design process. Femur—(a) load 3D model prediction, (b) add anterior–posterior cuts, (c) add medial–lateral cuts, (d) add pins, fillets, and chamfers. Tibia—(a) extract 2D profile from 3D model prediction, (b) extrude profile, (c) add bearing connection and posterior cuts, (d) add pin, fillets, and chamfers.

Figure 4

Distance heat maps for the femur (top row) and tibia (bottom row). (a) CT-3D model predictions and (b) custom components, both compared to ground truth models.

Figure 5

Calculation of the maximum OUH for the femur component (top row) and tibia plate (bottom row). (a) Implant components, (b) component edges, (c) component edges compared to ground truth geometry and identification of maximum OUH.

Figure 6

Box plot of pipeline results. Results split into three metrics: 3D model prediction RMSE, component RMSE and maximum OUH.