Knee Osteoarthritis Severity Grade Binary and Multiclass Classification Using ROI-Based Diverse Features

Abstract

Today, several diseaases have affected humans due to their unseen progression in early stages. Knee Osteoarthritis (OA) affecting most of the elderly mass is regarded as the common cause to restrict their normal activities and impairment. Therefore, a fast, speedy and cost effective early disgnosis of OA is the need of the hour. The work introduced in this article addresses the aforementioned issue and suggest an automated efficient diagnostic framework to classify OA severities based on X-ray or plain images. The Knee OA Severity Detetction Framework (K-OA-SDF) is used to study the impact on different class-based scenarios from binary to multiclass. The K-OA-SDF is subjected on X-ray images from Knee Osteoarthritis Dataset with Severity Grading (KODSG) of Kaggle dataset store to predict the knee OA grades. The K-OA-SDF is constructed using Region of Interest (ROI) Extraction Module, an efficient feature extraction module and a Convolutional Neural Network (CNN) based Classifier. The ROI module extract the significant portion of the X-ray images and enhances the contrast for beter quality features. The feature extraction module obtains diverse features from the ROI using image-based, object-based and traditional handcrafted features. Image-based features are acquired by dividing the ROI in two halves and averaging 16 columnar samples from both halves. Object-based blind features are obtained using modified VGG16 pre-trained network trained on the ImageNet dataset. Handcratfted features using quality descriptors are obtained to uplift the global and local details of the ROI. The CNN-based fine tuned classifier using 2000 (dimensionally reduced using Principal Component Analysis) features is able to outperform other state-of-the-art work in case of binary (98% accuracy) and ternay (88%) class categories while comparative results (71.17%) are obtained for five Kellegren Lawrence (KL) grades.