Recent studies have highlighted knee osteoarthritis as a notoriously incurable disease in the modern world. As the articular cartilage gradually wears out, it exacerbates the symptoms and impact on the affected area. Traditional methods of knee osteoarthritis identification primarily rely on knee x-rays, requiring professional judgment from doctors. However, the interpretation of these x-rays can be prone to human error, particularly when dealing with large volumes of scans and subtle variations within them.
In response to these challenges, convolutional neural network (CNN) based algorithms have emerged as a potential solution to automate parts of the diagnostic process. However, existing models often suffer from low test and validation accuracy, making them impractical for widespread use.
This paper proposes a more robust model for knee osteoarthritis diagnosis using transfer learning techniques. The aim is to develop a model capable of accurately identifying knee osteoarthritis even in the presence of minor modifications in the training data.
In this work, we utilized the Knee Osteoarthritis Dataset with KL Grading – 2018, sourced from Kaggle. This dataset is derived from the Osteoarthritis Initiative (OAI), a multi-center, longitudinal, prospective observational study focused on knee osteoarthritis (OA). The primary objective of the OAI is to identify biomarkers for OA onset and progression.
- Origin: The dataset is organized from the OAI.
- Age Range: The images included in the dataset correspond to individuals aged between 45 to 79 years.
- Labeling: The images in the dataset are categorized into five distinct classes based on the K&L Grading system:
- Label 0: Healthy knee images
- Label 1: Doubtful stage (osteophytic lipping) images
- Label 2: Minimal stage images
- Label 3: Moderate stage images
- Label 4: Severe stage images
- Sample Images: A visual representation of sample images from each category is provided in Figure 1.
Experiments were conducted on CNN-based state-of-the-art models by leveraging transfer learning techniques instead of a generic approach. Various architectures such as ResNet18, AlexNet, SqueezeNet, DenseNet, and EfficientNet were compared.
EfficientNet emerged as the top-performing model, outperforming other methods such as ResNet18, AlexNet, SqueezeNet, and DenseNet. The proposed model achieved a higher training accuracy of 93.8% and a validation accuracy of 92.3%. Furthermore, the model was trained to identify knee osteoarthritis into five classes according to the Kallgren Lawrence grading system.
| Model | Training Accuracy | Validation Accuracy | Training Loss | Validation Loss |
|---|---|---|---|---|
| Efficientnetb0 | 93.854% | 92.268% | 0.167 | 0.267 |
| DenseNet121 | 96.672% | 89.585% | 0.092 | 0.307 |
| Resnet18 | 96.389% | 92.163% | 0.087 | 0.288 |
| AlexNet | 93.521% | 89.073% | 0.174 | 0.365 |
| SqueezeNetV1 | 73.856% | 71.182% | 0.628 | 0.714 |
Best results are mentioned in bold characters
The results indicate that the proposed model not only achieves higher accuracy but also reduces loss compared to existing models. By leveraging transfer learning techniques, the model demonstrates improved robustness and effectiveness in knee osteoarthritis diagnosis, thus offering promising advancements in medical image analysis.
We extend our heartfelt appreciation to Prof.Taeg Keun Whangbo for his guidance and supervision throughout this project. His expertise and support were instrumental in steering the direction of our work and ensuring its success.
We would also like to acknowledge the contributions of the following individuals from the CT LAB:
- M.J. Aashik Rasool
- Dr.Shabir Ahmad
They played a key role in implementing the approach described in this project.