Development and validation of machine learning algorithms for early detection of ankylosing spondylitis using magnetic resonance images

BackgroundAnkylosing spondylitis (AS) is a chronic inflammatory disease affecting the sacroiliac joints and spine, often leading to disability if not diagnosed and treated early.ObjectiveIn this study, we present the development and validation of machine learning (ML) algorithms for AS detection only using Short Tau Inversion Recovery (STIR) sequenced magnetic resonance (MR) images.MethodsThe detection process is based on creating Gray Level Co-occurrence Matrices (GLCM) from MR images, followed by the computation of Haralick features and the training of ML-based models. A total of 696 MR images (AS+: 348, AS-: 348) were utilized for AS detection. Models were trained and tested on 70% of the dataset using a 10-fold cross-validation method to prevent overfitting, while the remaining 30% of the data was used for validation. In addition, care was taken to ensure that different images from the same patient were not split between the training and validation sets during this separation process to prevent potential data leakage.ResultsThe proposed ML-based model demonstrated superior performance during the validation phase (accuracy: 0.885, AUC: 0.941). The results of our study show promising outcomes when compared to previous works employing GLCM-based ML detection models. Conclusions: This study introduces a new perspective on AS detection, focusing on the assignment of ML techniques to STIR-sequenced MR images with a notable absence of literature on interpreting ML models for AS detection. This typology also addresses a lack of knowledge, as most models do not provide practical interpretability or knowledge alongside accurate prediction. The system also offers an effective strategy for early and correct diagnosis of AS, which is important for timely intervention and treatment planning.ResultsThe proposed ML-based model demonstrated superior performance during the validation phase (accuracy: 0.885, AUC: 0.941). The results of our study show promising outcomes when compared to previous works employing GLCM-based ML detection models. Conclusions: This study introduces a new perspective on AS detection, focusing on the assignment of ML techniques to STIR-sequenced MR images with a notable absence of literature on interpreting ML models for AS detection. This typology also addresses a lack of knowledge, as most models do not provide practical interpretability or knowledge alongside accurate prediction. The system also offers an effective strategy for early and correct diagnosis of AS, which is important for timely intervention and treatment planning.