Signal, Image and Video Processing Обработка сигналов
U-Net-mamba: a resource-efficient CNN-vision mamba network to segment meibomian glands from infrared meibography images
Аннотация
Meibomian Gland Dysfunction (MGD) is very common chronic eye disease caused when meibomian glands fails to produce enough oil or meibum to preserve tear film. The lack of meibum causes dryness in the eyelids affecting proper functioning of the eye which can cause irritation, vision blur, inflammation or burning. Segmenting meibomian glands from infrared images is a key approach for detecting dysfunctional glands. Several conventional deep learning methods, including U-Net and its variants, excel at capturing local spatial features but are inherently limited in modeling long-range dependencies across the glandular structure. To address this limitation, we propose U-Net-Mamba, a U-shaped hybrid architecture that integrates structured state space sequence modeling within a UNet framework for robust meibomian gland segmentation. The proposed model introduces a MambaConvBlock, which combines local convolutional feature extraction with Mamba-based state space modeling and channel-wise attention. The extracted spatial features are reshaped into sequence and the Mamba module is able to effectively capture global contextual relationships, while preserving spatial resolution through residual connections. The channel attention network preserve the discriminative gland features by adaptively adjusting the importance of informative channels. Experimental evaluation on MGD-1k dataset demonstrates that the proposed approach significantly outperforms existing U-Net-based and attention-based segmentation methods across multiple metrics, including Dice coefficient, Intersection over Union, accuracy, precision, sensitivity, specificity, HD95, and Boundary F1.