Brain Tumor Segmentation Meets Efficiency: Res-UNet Improved by Attention Mechanisms and Quantization

Brain tumor segmentation from Magnetic Resonance Imaging (MRI) images is a crucial step in medical diagnosis and treatment planning, which directly impacts clinical decision-making and patient outcomes, particularly in resource-constrained medical environments. However, achieving high segmentation accuracy while maintaining computational efficiency remains a challenge, particularly for complex tumor types. Therefore, the research aims to use the brain tumor segmentation dataset and the brain tumor MRI dataset from Kaggle to evaluate segmentation performance. The analysis also investigates the trade-off between model accuracy and efficiency by optimizing the Res-UNet architecture with attention mechanisms, including the Attention Gate (AG), Squeeze-and-Excitation (SE) Block, and the Convolutional Block Attention Module (CBAM). As the result, attention mechanisms improve feature representation and segmentation precision. Then, these procedures also add computational cost. To address this challenge, Dynamic Range Quantization (DRQ) compresses the model from 127 MB to 32 MB (75% reduction) and speeds up inference by 37% (0.3143 s to 0.1973 s). During the process, the best model, Res-UNet with AG, achieves a mean Intersection over Union (IoU) of 0.845 and drops only by less than 0.0004 after quantization. Unlike previous studies that explored attention or quantization in isolation, the researchers combine both to achieve accurate, efficient, and deployable brain tumor segmentation for resource-constrained settings.