Description
Artificial intelligence algorithms for automated assessment of limb alignment using long-leg X-rays
Introduction
The radiographic assessment of lower extremities is a standard protocol to estimate limb alignment, leg length discrepancies and joint orientation for the purpose of surgical planning as well as the post-operative follow-up. This task is time-consuming and challenging for non-specialized physicians, while artificial intelligence obtains excellent results in the analysis of medical images.
In this study, we present and evaluate a suite of deep learning algorithms to automatically determine the lower limb alignment angles and measure the prosthesis positioning on long-leg X-rays. And, we also assess the time savings achieved by orthopedics surgeons using AI algorithms.
Methods
We built a retrospective database containing more than 103,360 X-rays from 19,560 patients. We filtered this database using a set of quality control algorithms to obtain 2,200 AP long-leg X-rays.
Then, we trained several neural networks to measure prosthesis positioning (i.e. angle between the prosthesis components and the bones’ axes) and limb alignment using manually labeled anatomical landmarks. Therefore, we annotated landmarks on 1,000 images with 32 landmarks. The training procedure consisted of two steps: a first neural network determined the approximate localization of hip, knee and ankle; then we cropped each region of interest and optimized one independent network per region to produce precise landmarks.
Finally, we compared the angles calculated by our pipeline with the ones annotated. We also evaluated the median reading time for 2 orthopedics surgeons assisted by AI algorithms compared to manual readers using an open-source software without AI assistance.
Results
The algorithms automatically produce 6 angles for the left and right side of the patient: the Mechanical Femoral, Mechanical Tibial, Hip-Knee-Ankle, Hip-Knee-Shaft, Femoral Neck-Shaft and Joint Line Convergence.
We evaluated the performance of our algorithm by calculating the mean difference between the calculated and the measured angles. Our pipeline reached a mean angle error of 2.59° (std 3.94°). It should be noted that the alignment is measured both on pre- and post-operative X-rays.
The median reading time for orthopedics surgeons without and with IA assistance are respectively 240 (std 37) and 55 seconds (std 25). Substantial time saving to the order of 80% was observed for lower limb radiological measurements per patient for all readers by using AI algorithm.
Conclusions
To conclude, this study demonstrated that the use of artificial neural networks contributed to significant time savings in reliable radiological assessment of patients while keeping the same performance in the analysis of the images. It highlights the relevance of AI assistance for the assessment of long-leg X-rays. Their performance opens the way to a tool assisting in the precise, rapid and standardized measurements to integrate into the orthopedic surgeon’s routine.
