Classification of rheumatoid arthritis with a neural network
Rheumatoid arthritis (RA) is a chronic inflammatory joint disorder that can lead to destruction of joints, reduced quality of life, and increased mortality. Currently, the classification of RA is done with a system called Disease Activity Score 28 – C-reactive Protein. Its score depends on several subjective variables. Recently, a new method called OMERACT-EULAR Synovitis Score(OESS) has been suggested, which only uses ultrasound images for classification, and therefore has the potential to bring more objectivity to the treatment when coupled with the robotic platform of ROPCA for ultrasound scans. This paper investigates the possibility of using feature extraction with a convolutional neural network for classification of power Doppler ultrasound images of RA. Using InceptionV3, pre-trained on ImageNet, for extracting features, a self-constructed top was trained to classify RA with the OESS scoring system ranging from 0-3. An accuracy of 75.0 % was achieved for 4-class classification. An accuracy of 86.9 %, a sensitivity of 87.5 %, and a specificity of 86.4 % was achieved for binary classification. The dataset contained 1694 Doppler ultrasound images, 1342 used for training, 176 for validation, and 176 for test.