Diagnosis of different types of single suture craniosynostosis in infants with deep learning techniques

Abstract

Craniosynostosis (CSO) is characterized by premature fusion of skull sutures in infants. This early closure of one or more main sutures can lead to various skull and facial deformities and may cause developmental delay in children. Early diagnosis, crucial for effective treatment, traditionally relies on physical examination and 3D cranial imaging, which are often inaccurate or with the risk of X-ray exposure. This study presents a fully-automated deep learning-based method for diagnosing common types of single suture CSO using routine digital photographs of infants' heads. We employed a two-stage approach involving head segmentation and CSO type classification. First, mask region-based convolutional neural network (Mask R-CNN) was used for accurate head segmentation, achieving an average precision of 97.60 ​% and an average recall of 96.20 ​%. The segmented images were then classified into different CSO types using a modified VGG11 neural network. The classifier attained a training accuracy of 99.74 ​% and a test accuracy of 94.44 ​%, with high sensitivity and specificity for uni-coronal, metopic, and sagittal types. Our method illustrates high reliability and accuracy, offering non-invasive, accessible and accurate diagnostic instrument for early detection and patient screening.