LGSegNet: A lightweight multi-scale network for brain tumor segmentation from multi-modal MRI

Brain tumor segmentation from multi-modal magnetic resonance imaging is crucial for diagnosis, treatment planning, and monitoring. However, accurate segmentation remains challenging due to heterogeneous tumor morphology, ambiguous boundaries, and the need for computationally efficient models for practical deployment. To address these challenges, we proposed LGSegNet, a lightweight segmentation framework designed to improve both segmentation accuracy and efficiency. LGSegNet follows an encoder-decoder architecture, where an SSBlock-based encoder is employed to extract discriminative representations while reducing redundant feature responses. A Context Fusion Refinement Block is introduced at the bottleneck to enhance multi-scale spatial encoding and channel-wise discrimination. To further improve global semantic consistency, a Contextual Fusion Attention Block captures long-range dependencies. In the decoder stage, an attention-based SkipRefine module selectively filters skip connections to improve boundary delineation and suppress irrelevant activations. Experiments on two independent brain tumor datasets demonstrate the effectiveness of LGSegNet. On Dataset I, the LGSegNet achieves an average Dice score of 0.968 ± 0.161 on the training and 0.956 ± 0.158 on the validation across WT, TC, and ET, indicating stable generalization. On Dataset II, it attains an average Dice score of 0.977 ± 0.039 on training and 0.956 ± 0.028 on validation, demonstrating robust performance under diverse tumor appearances and imaging conditions. In addition to accuracy improvements, LGSegNet significantly reduces computational complexity, requiring only 6.4M parameters and 5.5 GFLOPs, with an inference time of approximately 0.2 s per sample. LGSegNet provides an efficient 2D slice-wise segmentation solution that support practical use after further validation in clinical and volumetric settings.