Performance analysis of Rheumatoid Arthritis using Convolutional Neural Networks

• Published in: NeuroQuantology Vol. 20; no. 9; p. 2817
• Main Authors: Mahesh Kumar A S, Mallikarjunaswamy, M S, Chandrashekara, S
• Format: Journal Article
• Language: English
• Published: Bornova Izmir NeuroQuantology 01.01.2022
• Subjects: Accuracy; Arthritis; Artificial neural networks; Automation; Data augmentation; Diagnosis; Fingers; Inflammation; Joints (anatomy); Magnetic resonance imaging; Rheumatoid arthritis; Wrist
• ISSN: 1303-5150
• DOI: 10.14704/nq.2022.20.9.NQ44328

Rheumatoid Arthritis (RA) is a kind of an autoimmune and chronic disease. RA generally observed with inflammation, swollen, stiffness, joint pain and loss of functionality in the joints. Inflammation starts at smaller joints of the body, in later stages inflammation spread to heart and other organs of the body. Initial symptom are shown to be less effective but in later stage it causes major difference in the functionality of the joints. Therefore, an accurate RA detection in its early stage is very much essential. Various modalities are being used for the purpose of RA diagnosis notably radiography, ultrasound and Magnetic Resonance Imaging (MRI). Even though various modalities used in the assessment of joint damage and position changes, plain radiography is the best and effective method. Different scoring methods used in the RA assessment but all scoring methods involved the joint evaluation of finger, hands, feet and wrist. The traditional scoring methods and manual diagnosis process require more human intervention and time. The development of CNN architecture for automated RA detection avoids manual method of preprocessing, handcrafted segmentation and classification. CNN architecture plays vital role in the RA disease diagnosis and automation. The work includes the development of four different CNN architectures for RA detection namely ResNet50, VGG16, DenseNet121 and InceptionV3. All the models have trained with augmented and non-augmented dataset. At the end of 50th epoch, InceptionV3 accuracy reached 98.91% with error of 1.85%. DenseNet121 model reached 97.27% with minimum error of 7.73%. InceptionV3 validation accuracy reached 98.8% on validation set that indicates that InceptionV3 has low variance compared to other models. Even inceptionV3 perform well on F1 score, precision, recall and specificity plots of other models