Accurate segmentation of lesions in demyelinating disease from brain MRI is necessary for disease monitoring, treatment planning, and outcome prediction. However, the task is incredibly challenging due to lesion heterogeneity, tiny lesion size, and low contrast with surrounding tissue. In this study, a multiscale residual V-Net with attention gating is proposed for automatic brain lesion segmentation. The network expands the standard V-Net framework by including multiscale residual blocks with receptive fields of different kernel sizes, allowing joint extraction of fine-grained lesion boundaries and broader contextual features. To further improve lesion localization, attention gates are integrated into skip connections to suppress irrelevant background activations and selectively highlight lesion-relevant features. For training, a hybrid loss function is introduced to balance volumetric sensitivity. In addition, a warmup cosine learning rate scheduler is employed to improve training stability and convergence. To ensure model interpretability, Grad-CAM–based explainable AI is further integrated, which visualizes the discriminative regions contributing to lesion segmentation, thereby providing transparency and helping radiological validation. The method was evaluated on the Laboratory of Imaging Technologies dataset, comparing against baseline U-Net and V-Net as well as a hybrid U-Net with ConvMixer architecture. Experimental results show that our model attains higher Dice coefficient (80.67), IoU (67.73), precision (80.03), and recall (81.42), particularly outstanding in the detection of white matter lesions. Our findings show that the integration of multiscale representation, attention modules, explainability, and adaptive optimization improves robustness in demyelinating lesion segmentation, providing a valuable tool for clinical research and patient monitoring.