Knee Osteoarthritis Severity Classification Using CNN and Image Enhancement Filters

Abstract

CNNs for the classification of knee osteoarthritis is a promising avenue to assist in clinical decision making. Our study examined the influence of several image enhancement filters on CNN performance for KOA severity classification using X-rays. The KOA dataset contained 8,260 labelled X-ray images of five KL grades (0-4) divided into three datasets (training, validation, and test). As we assessed the effects of the different image enhancement preprocessing filters on labelled X-ray images, we applied several typical image enhancement preprocessing filters such as the Sobel, Gaussian, CLAHE, Gabor, and Entropy filters in order to evaluate the effect they had on classification performance. The best classifier for KOA classification, using the CLAHE filter, had the best performance for the 5-layered CNN model with the best accuracy (85%), F1-score (0.72), and AUC score of (0.83). This exemplifies the usefulness of contrast enhancement in medical image classification. Findings from our study can demonstrate that image enhancement techniques contribute to the reproducibility of CNN based KOA grading systems and offer one more step towards more reliable and automated KOA assessment.