and thank you AOSSM and to my co-authors as well. So we know that posterior tibial slope has significant influences on knee biomechanics and kinematics, specifically with how it affects the tibial translation anteriorly, and its importance as a risk factor for failure after primary or revision anterior cruciate ligament reconstruction has been well-reviewed in the literature. There are various methods of measurement, so you could do this on 3D imaging, where you can measure either the lateral or the medial plateau. You can do this on 2D imaging. Do you use the mechanical axis? Do you use the anatomic axis? But I think we can all agree that the full-length standing knee to ankle is probably the best 2D modality to give you the most accurate measurement. However, when it comes to generalizability and availability, I think short-leg knee radiographs are probably the standard of care at most institutions. Our study question is how can we efficiently measure posterior tibial slope for a large cohort of patients with the available imaging modalities we have? And so as we've discussed, artificial intelligence has really flourished in the last several years, and it can be used to automate labor-intensive tasks, such as making these measurements on a large cohort of radiographs, and we hope to apply this subset of AI called deep learning computer vision, which allows for analysis of digitized images, digitized previously analog images by a computer software in ways that were previously thought impractical or impossible. Our purpose was therefore to develop a deep learning tool that can automatically measure posterior tibial slope on a given short-leg knee radiograph. We hoped to achieve a good concordance with measurements made by surgeons, which would define as a mean difference of plus or minus two degrees. And this is a diagram of our proposed workflow. So what we did was we took our cohort out of the Rochester Epidemiology Project. We had 390 radiographs of patients who underwent anterior cruciate ligament reconstruction. We balanced the two cohorts, we balanced the training and the testing data in terms of age, gender, and graft type. We then made manual segmentations by humans of the anatomic locations we thought would be important for our measurement of the tibial slope. And then we fed our manual segmentations to a deep learning algorithm, which then learned to identify the different anatomic areas and segment a new image. And once the deep learning algorithm achieved an appropriate threshold of learning in terms of accuracy of the segmentation, then we can utilize the algorithm, generate masks on a newly supplied image, and then we calculate our posterior tibial slope utilizing image processing, which is essentially just performing mathematical operations on the coordinates of pixels in the generated masks by the deep learning algorithm. We compared these to surgeon measurements, which were made by two independent reviewers based on previously described method as shown here. And our primary outcomes were, one, the performance of the segmentation algorithms. And we look at overlap of the segmentations made by the algorithm versus the manual annotations made by the surgeon. And we measure this utilizing something called Dice Similarity Coefficient. And a dice of one is basically perfect overlap of all the pixels. And then we looked at absolute mean difference in posterior tibial slope measurement between surgeon and machine. And then finally, we wanted to look at outliers, which we define here as absolute mean difference greater than or equal to five degrees. So looking at the performance of our segmentation algorithm, we achieved a dice on every segmentation of well above 0.8, which in the literature is generally the acceptable threshold for semantic segmentation model. When we looked at our results, we measured the difference between the two surgeon annotators, and there was a mean absolute difference of 1.3. And then when we looked at a comparison of surgeon versus machine, there was a mean absolute difference of 1.9, which was still within our goal of plus or minus two degrees. And looking at outliers, there were four images out of 90. So certainly the algorithm had the upside in terms of efficiency. This could complete measurement for 90 images in under a minute. We were able to package this into a digital application, although it can be deployed via the web, although we've not done so pending its external validation. So limitations, and I think this has been said before, AI is not kind of the end-all be-all, it's not a cure-all, so it's only as good as your best data. We very likely need to validate this with comparisons to measurements on the gold standard, which are long-leg radiographs, and then external data from other institutions. And then finally, it still remains to be seen whether or not these measurements really enable clinical prediction in terms of outcomes such as graph failure or revision. Thank you.

posterior tibial slope

knee biomechanics

anterior cruciate ligament reconstruction

measurement methods

deep learning tool