Breast tumor segmentation in ultrasound images presents a significant challenge due to the variability in tumor shapes and sizes. In this paper, we propose an enhanced model, Inception Vision Mamba U-Net (InViM-UNet), specifically designed to address these challenges. InViM-UNet integrates the Inception Module immediately after the patch embedding layer, enabling the model to effectively capture multi-scale features crucial for segmenting tumors with diverse characteristics. We evaluated the performance of InViM-UNet against several baseline models, including ViM-UNet, standard U-Net, MSVM-UNet, and ViT+UNet, on a breast ultrasound dataset. The results demonstrate that InViM-UNet achieves superior performance, with a Dice score of 0.7809, precision of 0.9240, F1 score of 0.7912, and IoU of 0.6889, outperforming all other models in breast tumor segmentation accuracy. While ViT+UNet exhibited higher Recall but lower Precision, and MSVM-UNet achieved high recall but a slightly lower precision, the standard U-Net showed the weakest performance. These findings highlight the effectiveness of the Inception Module at the input layer for capturing multi-scale features to improve segmentation outcomes for breast tumors, particularly those with diverse shapes and sizes. The proposed InViM-UNet model provides a promising solution to enhance the accuracy and reliability of breast ultrasound tumor segmentation.