Hybrid deep learning framework with feature fusion for cerebrospinal fluid leak detection

Early and accurate detection of Cerebrospinal Fluid (CSF) leakage is critical due to its life-threatening complications; however, manual evaluation of Magnetic Resonance (MR) scans is time-consuming and prone to diagnostic errors. Furthermore, the lack of specialized Decision Support Systems (DSS) for CSF leakage detection in the literature reveals a significant research gap. Current medical image segmentation and classification approaches often suffer from limited generalization capability, high computational cost, and inadequate modeling of long-range spatial dependencies. To overcome these challenges, this study proposes a novel two-stage hybrid deep learning framework for CSF leakage detection. In the first stage, ShallowMambaUNet, a lightweight yet effective segmentation model, integrates residual learning with Mamba-based state-space modeling to capture both local spatial details and global contextual dependencies. In the second stage, a multi-branch classification architecture combines EfficientNet, Mobile Vision Transformer (MobileViT), and Graph Convolutional Network (GCN) to extract complementary local, global, and relational features. The inclusion of segmentation outputs in the classification stage further improved performance. Experimental results demonstrate that the proposed model achieved a Dice score of 0.9779 and an IoU score of 0.9568 in segmentation, while outperforming existing methods in classification with 98.21% accuracy, a 98.25% F1 score, 96.5% MCC, 96.4% Kappa, 98.2% G-mean, and a 0.99 AUC. Furthermore, evaluations on multiple publicly available brain magnetic resonance datasets confirm the strong generalization capability of the proposed framework.