Abstract
OBJECTIVES: Sturge-Weber syndrome (SWS) is a congenital neurological disorder occurring in the early childhood. Timely diagnosis of SWS is essential for proper medical intervention that prevents the development of various neurological issues. Leptomeningeal angiomas (LA) are the clinical manifestation of SWS. Detection of LA is currently performed by manual inspection of the magnetic resonance images (MRI) by experienced neurologist, which is time-consuming and lack of inter-rater consistency. The aim of the present study is to investigate automated LA detection in MRI of SWS patients. METHODS: A Mamba-based encoder-decoder architecture was employed in the present study. Particularly, a multi-scale multi-scan strategy was proposed to convert 3-D volume into 1-D sequence, enabling capturing long-range dependency with reduced computation complexity. Our dataset consists of 40 SWS patients with T1-enhanced MRI. The proposed model was first pre-trained on a public brain tumor segmentation (BraTS) dataset and then fine-tuned and tested on the SWS dataset using 5-fold cross validation. RESULTS AND CONCLUSION: Our results show excellent performance of the proposed method, e.g., Dice score of 91.53% and 78.67% for BraTS and SWS, respectively, outperforming several state-of-the-art methods as well as two neurologists. Mamba-based deep learning method can automatically identify LA in MRI images, enabling automated SWS diagnosis in clinical settings.