LBMNet: a hybrid multi-scale CNN-Mamba framework for enhanced 3D stroke lesion segmentation in MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Stroke is a leading cause of death and disability globally.
• Accurate segmentation of brain stroke lesions from MRI is crucial for diagnosis and treatment.
• Existing methods face challenges with lesion variability and detecting small lesions.

Purpose of the Study:

• To propose LBMNet, a novel CNN-Mamba network for accurate brain stroke lesion segmentation.
• To address limitations of current methods in handling diverse lesion sizes and morphologies.
• To improve the detection of small stroke lesions.

Main Methods:

• Developed LBMNet, a hybrid CNN-Mamba network integrating multi-scale convolutional encoding and Mamba-based decoding.
• Employed a top-down LSC module in the encoder for cross-scale representations.
• Designed a Bidirectional Spatial Context Mamba (BSC-Mamba) decoder with adaptive spatial convolutions and asymmetric adaptive gated feature fusion (BAGF).

Main Results:

• Achieved state-of-the-art performance on benchmark datasets (Dice: 67.57% on ATLAS v2.0, 82.03% on ISLES 2022).
• Demonstrated significant improvements in segmenting small brain stroke lesions compared to existing models.
• LBMNet showed robust and efficient performance across varied lesion characteristics.

Conclusions:

• LBMNet offers a robust and efficient framework for brain stroke lesion segmentation.
• The proposed method shows strong clinical potential for improving patient care.
• LBMNet effectively handles the heterogeneity of stroke lesions, particularly small ones.