Deep learning-based craniosynostosis classification via suture segmentation and mask-weighted classification.
Authors
Affiliations (7)
Affiliations (7)
- Department of Electronic Engineering, Kwangwoon University, Seoul, Republic of Korea.
- Department of Plastic and Reconstructive Surgery, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Graduate School of Smart Convergence, Kwangwoon University, Seoul, Republic of Korea.
- Department of Plastic and Reconstructive Surgery, Kyungpook National University School of Medicine, Daegu, Republic of Korea.
- Department of Plastic and Reconstructive Surgery, Jeonbuk National University Medical School, Jeonju, Republic of Korea.
- Research Institute of Clinical Medicine, Biomedical Research Institute, Jeonbuk National University Hospital, Jeonju, Republic of Korea.
- Department of Orthodontics, Dental Research Institute, Seoul National University School of Dentistry, Seoul, Republic of Korea.
Abstract
Early diagnosis of craniosynostosis (CSO) is critical to preventing neurological complications, yet skull X-ray interpretation remains subjective, and existing deep learning models often rely on secondary cranial deformations rather than the primary pathology. To address this limitation, we propose an Integrated Suture Segmentation and Classification Pipeline that explicitly learns suture information to enhance anatomical validity and diagnostic accuracy. We constructed a balanced dataset of 1,088 skull X-ray images from 368 unique patients and developed a segmentation model to identify coronal, sagittal, and lambdoid sutures. Crucially, we introduced a Mask-weighted 4-channel Input strategy, utilizing predicted suture probability maps as weights to guide the classification model's attention toward suture regions. Experimental results demonstrated that the proposed method with a DenseNet-161 backbone achieved an image-level Accuracy of 0.925 and an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.980. Furthermore, exam-level diagnosis via multi-view aggregation significantly improved performance, yielding an Accuracy of 0.941 and an AUROC of 0.994. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis demonstrated that the model's attention is primarily directed toward specific suture lines rather than global skull shape, suggesting that the model prioritizes anatomical features over secondary deformations commonly seen in conditions like positional plagiocephaly. This study presents a clinically interpretable and high-performance deep learning framework, highlighting its potential as a robust computer-aided referral decision support tool for primary care settings, facilitating timely specialist assessment while minimizing the need for unnecessary radiation-intensive CT scans.