LF-MAMBA: A State Space Model-Based Framework for Accurate Bilateral Hippocampi Segmentation in Low-Field Pediatric MRI

Precise segmentation of bilateral hippocampi in pediatric brain MRI is essential for studying early brain development and diagnosing neurodevelopmental disorders. Current deep learning methods face challenges with low-field MRI due to lower image quality, signal-to-noise ratios, and tissue contrast. We introduce LF-Mamba, a UNet-based model specifically designed for low-field pediatric MRI segmentation, integrating structured state-space modeling for long-range dependency capture and frequency domain attention for effective feature fusion. We extensively evaluated LF-Mamba against state-of-the-art models on the LISA dataset, with results demonstrating that LF-Mamba outperforms existing models, achieving a mean Intersection over Union (mIoU) of 45.23% and a Dice Similarity Coefficient (DSC) of 60.24%. These results indicate that LF-Mamba enhances segmentation performance in low-field MRI, supporting early diagnosis and intervention for neurodevelopmental disorders in resource-limited settings.