2-Stage classification of knee joint thermograms for rheumatoid arthritis prediction in subclinical inflammation

Presence of inflammation in knee joint is the early indication of arthritis. In this paper, we performed the inflamed region segmentation from knee joint thermograms for structural feature extraction based knee abnormality prediction. Existing four popular segmentation techniques are investigated, namely K-means, Fuzzy C-means, Otsu, Single seeded region growing. We proposed modified multi-seeded region growing method that generates 98.6% accurate segmentation rate compared to ground truth of inflammation. Based on the spread of the inflammation oriented structural feature analysis, in the first stage of classification we classified arthritis affected knee joint thermograms, and all other types of thermograms (non-arthritis) with 91% accuracy. Among different types of arthritis, the most damaging type that causes disability of joints in long run is known as rheumatoid arthritis (RA). Early diagnosis of RA in subclinical stage enormously helps clinicians to decrease the disease affect. In second stage of classification, we integrated the RA and non-RA categorization by extracting texture, shape and frequency level features. Experiment shows that the combination of all features decreases the accurate detection rate of RA classification. To increase the classification rate, we incorporated the accuracy based feature selection procedure. The RA classification rate obtained with accuracy based feature selection is 73% whereas existing support vector machine-recursive feature elimination (SVM-RFE) and RELIEF methods provide 67% and 71% correct classification rate respectively. The area under the curve (AUC) of accuracy based feature selection, SVM-RFE, and RELIEF for RA classification are 0.72, 0.65 and 0.67, respectively and it indicates better classification outcome of the accuracy based feature selection method.