ConvNeXt Based Deep Learning Framework for Knee Osteoarthritis Classification from Infrared Thermographic Images

Abstract

Knee osteoarthritis (KOA) is a progressive musculoskeletal disorder that affects physical movement and quality of life. Infrared thermography (IRT) provides a non-contact and non-ionizing imaging modality capable of capturing the spatial variations in the skin-surface temperature associated with the physiological changes around the knee. This paper proposes a deep learning framework for five-class KOA severity classification using thermal images from the publicly available KOA dataset. The thermal images will be preprocessed, anatomical region-of-interest extraction, and intensity normalization before model development. ConvNeXt-Tiny classification model is adopted as the principal deep-learning backbone for experimental protocol. The proposed model was evaluated on 718 test thermal images across five KOA severity categories and model achieved a balanced accuracy of 90.98%, with macro-averaged precision, recall, and F1-score of 91.01%, 90.98%, and 90.99%, respectively. The Matthews correlation coefficient and Cohen's kappa were both 0.887, indicating strong agreement between the predicted and reference severity categories. The proposed framework is designed to learn discriminative thermal representations associated with progressive KOA severity while providing interpretable evidence of model decision-making.