Analysis of CNN and Quantized CNN Model’s Performance for Osteoarthritis Identification

Abstract

Aging populations have made osteoarthritis (OA) a serious worldwide health problem. Radiography, Computed Tomography (CT), ultrasonic, Magnetic Resonance Imaging (MRI), and thermal imaging are all utilized to discover and diagnose OA. Early OA may be detected by non-invasive, radiation-free, quick, and accurate infrared (IR) thermal imaging. OA screening may benefit from IR thermography since it gives useful details on the temperature and vascular status of the joints.  Thermal imaging for the detection of OA disorders sometimes requires hand analysis and interpretation, which strongly relies on the clinical expertise of the examining physician. The goal of the paper is to automate the OA screening using IR thermographic images. In this study, the required dataset is collected using the IR imaging method. The collected dataset consists of normal and arthritis-affected thermal images. The dataset is then prepared to be compatible with the Convolution Neural Network (CNN) of the Deep Learning (DL) model and statistical parameters such as mean, mode, mode, kurtosis, etc. are derived and the correlation between the parameters is drawn using the covariance matrix. The dataset is then visualized using graphical plots to view the distribution of the statistical parameters. The dataset is then pre-processed using the normalization method and then analyzed using the CNN model. To find the efficiency of the DL model performance metrics such as accuracy, precision, loss, F-1 score, recall, etc. are calculated. To reduce the complexity and to reduce the computational time, the dataset is quantized using the optimization method and a comparison between the trained model and the quantized model is drawn.