Machine learning for detecting moyamoya disease in plain skull radiography using a convolutional neural network

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Moyamoya disease (MMD) diagnosis often relies on complex imaging, but morphological differences in MMD patients' skulls offer potential for simpler identification.
• Deep learning (DL) advancements present an opportunity to explore automated MMD detection using readily available skull radiographs.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL) algorithms in distinguishing moyamoya disease (MMD) from control cases using plain skull radiograph images.
• To evaluate the diagnostic performance of DL models in identifying MMD based on craniofacial morphology.

Main Methods:

• A dataset of 345 MMD skull images (ages 18-50) and 408 control trauma patient images was curated and split into 70% training and 30% testing sets.
• A six-convolution layer deep learning model was trained and evaluated using accuracy, sensitivity, specificity, and AUROC. Gradient-weighted class activation mapping (Grad-CAM) was used for visualization, and external validation was performed.

Main Results:

• The DL model achieved 84.1% accuracy, 0.84 sensitivity, 0.84 specificity, and an AUROC of 0.91 on the institutional test set.
• Gradient-weighted class activation mapping indicated that the model primarily focused on the viscerocranium (lower face) when identifying MMD.
• External validation on a separate hospital's dataset yielded an overall accuracy of 75.9%.

Conclusions:

• Deep learning demonstrates considerable accuracy in differentiating moyamoya disease (MMD) from controls in specific age groups using plain skull radiographs.
• The findings suggest that craniofacial features, particularly those of the viscerocranium, may serve as important indicators for MMD.
• This study highlights the potential of AI-powered analysis of skull X-rays for MMD diagnosis and further research into MMD-related morphological changes.