An improved activation function for the recognition of knee osteoarthritis severity

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Traditional knee osteoarthritis (KOA) classification methods are computationally intensive and inefficient for real-time diagnosis.
• Existing models struggle with performance limitations, especially on small datasets.
• There is a need for efficient and accurate KOA severity recognition models.

Purpose of the Study:

• To develop a deep transfer learning approach for KOA severity recognition without object detection.
• To improve classification performance on small sample datasets.
• To optimize computational resources for KOA diagnosis.

Main Methods:

• Utilized a dataset of 3300 digital X-ray images of KOA patients.
• Employed a transfer learning approach with a pretrained Inception-v3 backbone.
• Introduced a novel activation function (AvRELU) and used Bayesian optimization for model fine-tuning and classifier construction.

Main Results:

• Achieved 95% accuracy, 95% F1-score, and 93% kappa using five-fold cross-validation.
• Demonstrated superior performance in KOA severity recognition.
• Validated statistical significance using Kendall's tau-b correlation analysis.

Conclusions:

• The proposed method significantly enhances model performance on small-sample KOA datasets.
• The model maintains lightweight and low-resource characteristics, suitable for practical applications.
• This approach outperforms most existing KOA severity recognition methods.