The Knee · 2026
An improved activation function for the recognition of knee osteoarthritis severity.
СтатьяНазвание и аннотация: Перевод в очереди
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.
Chang Shuaishuai, Duan Hongliang, Wu QingE
Аннотация
BACKGROUND: Existing knee osteoarthritis (KOA) severity classification methods typically rely on a combination of object detection algorithms and classification algorithms. However, this approach not only increases the computational burden and time costs but also reduces the efficiency of real-time diagnosis, which makes it difficult to meet the needs of practical applications. To address the performance limitations of KOA severity recognition models that operate without target detection algorithms, a deep transfer learning approach incorporating a novel activation function (AvRELU) was proposed. The goal was to improve classification performance, particularly for small sample datasets, while optimizing computational resources.
METHODS: A dataset consisting of 3300 digital X-ray images of KOA patients was utilized. During model training and evaluation, five-fold cross-validation was used to assess robustness, and the dataset was divided into training, validation, and testing sets through a stratified sampling method at an 8:1:1 ratio. A transfer learning approach utilizing a pretrained Inception-v3 backbone was proposed, where Bayesian optimization automated both the fine-tuning process and downstream classifier construction. To improve the model performance, the activation function AvRELU was introduced in the network layer of the downstream model. Moreover, Kendall's tau-b correlation analysis was employed to evaluate the statistical significance of differences in the predicted Kellgren-Lawrence grades among the different methods.
RESULTS: Using five-fold cross-validation, the proposed method achieved an average test set performance of 95% accuracy, 95% F1-score, and 93% kappa. These results demonstrated the superior performance of the method in KOA severity recognition.
CONCLUSIONS: The method proposed here not only significantly improves model performance on small-sample datasets but also maintains the model's lightweight and low-resource characteristics. Moreover, it achieves better performance in KOA severity recognition than most existing methods.
METHODS: A dataset consisting of 3300 digital X-ray images of KOA patients was utilized. During model training and evaluation, five-fold cross-validation was used to assess robustness, and the dataset was divided into training, validation, and testing sets through a stratified sampling method at an 8:1:1 ratio. A transfer learning approach utilizing a pretrained Inception-v3 backbone was proposed, where Bayesian optimization automated both the fine-tuning process and downstream classifier construction. To improve the model performance, the activation function AvRELU was introduced in the network layer of the downstream model. Moreover, Kendall's tau-b correlation analysis was employed to evaluate the statistical significance of differences in the predicted Kellgren-Lawrence grades among the different methods.
RESULTS: Using five-fold cross-validation, the proposed method achieved an average test set performance of 95% accuracy, 95% F1-score, and 93% kappa. These results demonstrated the superior performance of the method in KOA severity recognition.
CONCLUSIONS: The method proposed here not only significantly improves model performance on small-sample datasets but also maintains the model's lightweight and low-resource characteristics. Moreover, it achieves better performance in KOA severity recognition than most existing methods.