LGG Segmentation Dataset.

Deep learning has emerged as the preeminent technique for semantic segmentation of brain MRI tumors. However, existing methods often rely on hierarchical downsampling to generate multi-scale feature maps, effectively capturing fine-grained global features but struggling with large-scale local features due to insufficient network depth. This limitation is particularly detrimental for segmenting diminutive targets such as brain tumors, where local feature extraction is crucial. Augmenting network depth to address this issue leads to excessive parameter counts, incompatible with resource-constrained devices. To tackle this challenge, we propose that object recognition should exhibit scale invariance, so we introduce a shared CNN network architecture for image encoding. The input MRI image is directly downsampled into three scales, with a shared 10-layer convolutional network employed across all scales to extract features. This approach enhances the network’s ability to capture large-scale local features without increasing the total parameter count. Further, we utilize a Transformer on the smallest scale to extract global features. The decoding stage follows the UNet structure, incorporating incremental upsampling and feature fusion from previous scales. Comparative experiments on the LGG Segmentation Dataset and BraTS21 dataset demonstrate that our proposed LiteMRINet achieves higher segmentation accuracy while significantly reducing parameter count. This makes our approach particularly advantageous for devices with limited memory resources. Our code is available at https://github.com/chinaericy/MRINet.