Automated diagnosis of Sturge-Weber syndrome by detecting leptomeningeal angiomatosis from MRI using a multi-task deep learning framework.
Authors
Affiliations (6)
Affiliations (6)
- Department of Pediatric Neurology, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- School of Information Science and Technology, ShanghaiTech University, Shanghai, China.
- Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands.
- Faculty of Applied Science, Macao Polytechnic University, Macau, Macao SAR, China.
- Department of Pediatric Neurosurgery, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- State Key Laboratory of Advanced Medical Materials and Devices, Shanghai, China.
Abstract
Sturge-Weber Syndrome (SWS) is a rare neurocutaneous disorder requiring early and accurate diagnosis. This study aims to develop a deep learning framework for automatic diagnosis of SWS by detecting leptomeningeal angiomatosis (LA) in brain magnetic resonance images (MRI). This retrospective study includes 40 SWS patients and 101 healthy controls. T1-weighted MRI were collected over 15 years with different scanners. Multi-scale harmonization was proposed to normalize the images to uniform mean and standard deviation. We developed a deep learning model based on the UNet framework with convolutional neural networks and transformer. A multi-task pipeline was designed to perform LA segmentation and SWS classification. The model was pre-trained on a public dataset and fine-tuned and tested on our dataset using a 5-fold cross validation. The segmentation and classification results were compared with the ground truth and human readers using various metrics at the voxel and LA level. For LA segmentation, our model achieved overall Dice, voxel-level sensitivity, specificity, accuracy, and kappa of 0.768, 0.772, 0.992, 0.983, and 0.760, respectively. The LA volume segmented by our model showed a high agreement and consistency with the ground truth, outperforming the human readers. For SWS classification, our model achieved sensitivity, specificity, accuracy, and AUC of 0.974, 1.000, 0.993, and 0.998, respectively. We developed a multi-task learning framework for automated LA segmentation and SWS classification. This single-center proof-of-concept study shows promising performance in both tasks. External validation may boost the clinical application of the developed model, assisting less-experienced neurologists for improved SWS diagnosis.