A fully automatic fine tuned deep learning model for knee osteoarthritis detection and progression analysis

Knee osteoarthritis (KOA) sufferers have one of the highest disability-adjusted life years. The entire knee
joint is affected by KOA. KOA is a condition that makes it hard for the knee to move normally. In KOA, the
damage to the joints is irreversible, and the only treatment is a total knee replacement (TKR), which is
expensive and only lasts a short time, especially for obese people. The individual’s social isolation and
low quality of life are significant outcomes of KOA. Despite being time-consuming and highly subject
to user variation, segmentation, manual diagnosis, and annotation of knee joints are still the most common procedure used in clinical practices to diagnose osteoarthritis. To overcome the above limitations of
the previously mentioned widely used procedure and reduce diagnostic errors made by doctors, we proposed a fine-tuning KOA diagnosis model using the DenseNet169 deep learning (DL) technique to
improve the efficiency of KOA diagnosis. Our model will reduce the cost of diagnosis, speed up diagnosis,
and delay disease progression, enhancing the procedure from the patient’s perspective. The proposed
model will determine the degree of KOA diseases by making multi-classification and binary classifications of the KOA severity. It will successfully localize the opacities’ peripheral, diffuse distribution, and
vascular thickening. Therefore, the proposed model would enable clinicians to understand the primary
causes of KOA through this localization. We evaluated the proposed model over the OAI dataset. The
OAI dataset was pre-processed by artifact removal, resizing, contrast handling, and a normalization technique. The proposed model was evaluated and compared with recent classifiers. In multi-classification,
the DenseNet169 model achieved 95.93%, 88.77%, 95.41%, 85.8%, and 87.08% for accuracy, sensitivity,
specificity, precision, and F1-score, respectively. In binary classification, the accuracy, sensitivity, specificity, precision, and the F1-score of the DenseNet169 model were 93.78%, 91.29%, 91.29%, 87.57%, and
89.27%, respectively. Therefore, the proposed model is an unrivaled perceptive outcome with tuning as
opposed to other ongoing existing frameworks