Designing of an inflammatory knee joint thermogram dataset for arthritis classification using deep convolution neural network.

ABSTRACT Limited application of thermography for inflammatory joint disease diagnosis is due to unavailability of joint thermogram dataset and formulated protocol of data acquisition. Focusing on the limitations, we aimed on creation and analysis of knee thermogram dataset by introducing standardized protocols of acquisition. The dataset named as “Infrared Knee Joint Dataset”, and includes healthy, and three different types of arthritis affected knee thermograms. Dataset validation and inflammation oriented ground truth generation procedures are also mentioned in this study. After data acquisition, thermograms are preprocessed and segmented. Finally, the system separates healthy and abnormal knee thermograms, and classifies those abnormal thermograms into three classes. For the classification, conventional feature-based techniques combined with shallow learning as well as deep learning have been used. The experimental results show the following: 1) classification of healthy and arthritis affected knee thermogram achieved 92% accuracy with SVM and 96% using VGG19; 2) In inter-arthritis classification VGG16 has shown the highest accuracy of 86% through ROI-based classification. Creation of standardized knee thermogram dataset and application of deep learning methodology diagnosis arthritis-oriented knee abnormality non-invasively. The described database acquisition protocol and classification strategies could contribute to the designing of accurate and robust image-based arthritis diagnosis systems.