LBMNet is a hybrid CNN-Mamba framework designed for improved 3D segmentation of stroke lesions in MRI. Stroke is a major cause of death and disability, and accurate segmentation of lesions is crucial for diagnosis and treatment. Existing methods fail with high variability in lesion size and shape, especially for small lesions due to the limitations of CNNs and transformers. The LSC encoder module captures representations at different scales from top to bottom. The BSC-Mamba decoder integrates two-way state space modeling with adaptive spatial convolutions for global dependencies with linear complexity. Asymmetric adaptive gated feature fusion (BAGF) combines encoder and decoder functions selectively. Experiments on the 302022LES and ATLAS2.2.2 datasets achieved a Dice coefficient of 67.57% and significant improvements in small lesion segmentation over CNN, Transformer and hybrid models. The framework has strong clinical potential for a variety of lesion sizes and morphologies.