A New Approach to Classify Knee Osteoarthritis Severity from Radiographic Images based on CNN-LSTM Method

Abstract

This paper introduces a new approach to quantify knee osteoarthritis (OA) severity using radiographic (X-ray) images. Our new approach combines preprocessing, Convolutional Neural Network (CNN) as a feature extraction method, followed by Long Short-Term Memory (LSTM) as a classification method. Preprocessing is conducted by manually cropping on the knee joint with dimensions of 400 x 100 pixels. The public dataset used to evaluate our approach is the Osteoarthritis Initiative (OAI) with very promising results from the existing approach where this dataset has information about the KL grade assessment for both knees (right and left). OAI is a multicenter and prospective observational study of knee OA. The purpose of this dataset is to develop public domain research resources to facilitate scientific evaluation of biomarkers for OA as a potential replacement endpoint for disease development. We have experimented by using three-fold cross-validation, where the first 2/3 data becomes the training data, while the last 1/3 data work as the testing data. Those groups data are being rotated with no overlap. Obtained results demonstrate that the mean accuracy is 75.28 %, and the mean loss function using cross-entropy is 0.09. These results outperform the deep learning methods that have been implemented before.