A comparative study of deep learning models for the classification of knee osteoarthritis in X-ray images

• Main Authors: Faris Jamil, Muhammad Irfan, Samad, Rosdiyana, Pebrianti, Dwi, Mustafa, Mahfuzah, Abdullah, Nor Rul Hasma, Noordin, Nurul Hazlina
• Format: Proceeding Paper
• Language: English; English
• Published: IEEE 2024
• Subjects: T Technology (General)
• Online Access: http://irep.iium.edu.my/115355/1/115355_A%20comparative%20study%20of%20deep%20learning%20models.pdf; http://irep.iium.edu.my/115355/7/115355_A%20comparative%20study%20of%20deep%20learning%20models_SCOPUS.pdf; http://irep.iium.edu.my/115355/; https://ieeexplore.ieee.org/document/10652557
• Institution: Universiti Islam Antarabangsa Malaysia

Knee osteoarthritis (OA) is a prevalent degenerative j oint disease that causes significant pain and disability. Early and accurate diagnosis is crucial for effective management, traditionally relying on manual assessment of X-ray images by medical experts. This manual approach can be time-consuming and subject to variability. Recent advancements in deep learning offer promising automated solutions for knee OA diagnosis from X-ray images. This study evaluates and compares the performance of three state-of-the-art deep learning models - Convolutional Neural Network (CNN), ResNet-50, and DenseNet-121 - for the classification of knee osteoarthritis using X-ray images. A comprehensive dataset of 10,930 knee X-ray images, labeled as normal or osteoarthritis, was utilized for training and testing the models. Rigorous hyperparameter tuning and optimization were performed to achieve optimal performance. The results demonstrate that the CNN model consistently outperformed the ResNet-50 and DenseNet-121 models across various evaluation metrics, including overall accuracy (97.01%), precision (95.94%), recall (98.97%), specificity (95.85%), and F1 score (96.98%). The CNN model exhibited superior capability in distinguishing between normal and osteoarthritis knees, attributed to its ability to effectively learn hierarchical features and its relatively simpler architecture.