Haralick Feature-Based Deep Learning Model for Ankylosing Spondylitis Classification Using Magnetic Resonance Images

Ankylosing spondylitis (AS) is a chronic inflammatory disease affecting the sacroiliac joints and spine, causing pain, stiffness, and structural damage. The diagnosis of Ankylosing spondylitis (AS), despite its global prevalence of $0.1 \%$ to $1.4 \%$, is often delayed by 5 to 6 years, highlighting the need for improved assessment tools. While previous studies focused on deep learning (DL) models trained on image data, our study introduces a novel approach. We utilize grey-level co-occurrence matrices (GLCM) derived from magnetic resonance images with our DL-based AS classification system trained on 21 Haralick features extracted from GLCM. This method provides insights into surface texture characteristics, potentially streamlining the classification process. Our model, trained on numerical data instead of images, demonstrates relatively high accuracy and discriminatory power, offering a promising alternative to computationally intensive training methods. Moreover, on the validation dataset, the model achieves $92.8 \%$ accuracy, $90.6 \%$ sensitivity, F1-Score of 0,939 and an area under the curve (AUC) of 0.9417. These findings indicate that numerical inputs derived from Haralick Features of magnetic resonance images can be used to train deep learning decision-support systems, which might be used to diagnose AS associated with delayed diagnoses and worse patient quality of life.