The Journal of arthroplasty · 2025
TKA-AID: An Uncertainty-Aware Deep Learning Classifier to Identify Total Knee Arthroplasty Implants.
СтатьяПоддержка исследования: вне правительства СШАНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Mulford Kellen L, Saniei Sami, Kaji Elizabeth S, Grove Austin F, Girod-Hoffman Miguel, Rouzrokh Pouria, Abdel Matthew P, Taunton Michael J, Wyles Cody C
Аннотация
BACKGROUND: A drastic increase in the volume of primary total knee arthroplasties (TKAs) performed nationwide will inevitably lead to higher volumes of revision TKAs in which the primary knee implant must be removed. An important step in preoperative planning for revision TKA is implant identification, which is time-consuming and difficult even for experienced surgeons. We sought to develop a deep learning algorithm to automatically identify the most common models of primary TKA implants.
METHODS: We used an institutional total joint registry to pull images and implant data for 9,651 patients (N = 111,519 images). We trained a deep learning model based on the EfficientNet architecture to identify nine different TKA systems across all common knee radiographic views. Model performance was assessed on internal held-out test set and external test set. Conformal prediction was employed to provide uncertainty estimates, and an outlier detection system alerts the user if an image is potentially outside of the model's trained expertise.
RESULTS: The average model accuracy on the held-out test set was 99.7%. The outlier detection system flagged 93% of images in the test set which were marked as outliers during a manual clean of the dataset. On the external test set, the model made one error out of 301 images. The model can process approximately 30 images/second.
CONCLUSION: We developed an automated knee implant identification tool that can classify nine different implant designs. Importantly, it works on multiple radiographic views and utilizes uncertainty quantification and outlier detection as safety mechanisms.
METHODS: We used an institutional total joint registry to pull images and implant data for 9,651 patients (N = 111,519 images). We trained a deep learning model based on the EfficientNet architecture to identify nine different TKA systems across all common knee radiographic views. Model performance was assessed on internal held-out test set and external test set. Conformal prediction was employed to provide uncertainty estimates, and an outlier detection system alerts the user if an image is potentially outside of the model's trained expertise.
RESULTS: The average model accuracy on the held-out test set was 99.7%. The outlier detection system flagged 93% of images in the test set which were marked as outliers during a manual clean of the dataset. On the external test set, the model made one error out of 301 images. The model can process approximately 30 images/second.
CONCLUSION: We developed an automated knee implant identification tool that can classify nine different implant designs. Importantly, it works on multiple radiographic views and utilizes uncertainty quantification and outlier detection as safety mechanisms.