Knee arthroplasty is now the most common method of treating the end stages of knee wear. In order to improve operative treatment, new approaches to the placement of both femoral and tibial components of knee arthroplasty have emerged in the last ten years. These new approaches are based on the premise that a kinematically placed prosthesis should better follow the normal motion of the knee joint. Therefore, the accuracy of prosthesis placement is becoming more and more important when performing knee arthroplasty.
Currently, the arbitrarily defined tolerance limit for the deviation of the placement precision from the plan precision for a given angle is $\pm$3$^{\circ}$. Pre-operative angle measurements are routinely taken on long-axis X-rays, which represent a frontal 2D projection of the lower limb. This thesis investigates the accuracy of automated measurements on long-axis X-rays of the lower limb using the YOLO model compared to manual (reference) measurements by two experts in radiology and orthopedics. The acceptable deviation of the automated X-ray measurements was empirically determined to be $\pm$3 mm or $\pm$3$^{\circ}$ relative to the reference measurements.
The experiments systematically evaluated the proposed automatic method for the localization of the landmark by investigating the impact of (i) hyperparameter tuning, (ii) different model sizes, (iii) input image sizes and (iv) post-processing methods such as the principal component analysis and the Canny edge detector, for final landmark location refinement. The experimental results showed that the use of a YOLO model with 11,6 million parameters, labelled $"$small$"$, using weight transfer learning and the random gradient descent optimization method, the use of enrichment with rigid geometric mappings, an input image size of 1920$\times$480, and post-processing techniques (principal component method and Canny edge detector) gave the best results.
A dataset of 188 long-axis X-rays of 107 patients was collected. Two experts manually annotated the locations of the anatomical landmarks on these images. The average deviation between the reference annotations of the two experts was 1,99 mm on the X-axis and 1,22 mm on the Y-axis of the X-rays. When using the YOLO model, the average deviation from the reference dataset was 2,44 mm on the X-axis and 2,63 mm on the Y-axis, an euclidean distance of 3.98 mm, an average absolute deviation of the predicted HKA angle from the reference HKA angle of 0.57$^{\circ}$, and method's success rate of 97.29\% on the test dataset (37/38 images), which meets the predefined criteria for acceptable performance level.
Automatic localization of anatomical landmarks on long-axis X-rays using the YOLO model with postprocessing achieves sufficient accuracy compared to manual annotations by experts. We anticipate that with further parameter tuning, larger models with more parameters, better post-processing and larger datasets, the accuracy of the measurements could be further improved.
|
