Optimizing Brain Tumor Segmentation: Comparing Smoothing Techniques for Network Performance

In the field of brain tumor segmentation, achieving high accuracy with lightweight and fast networks remains a major challenge. This paper describes a novel strategy for developing an efficient and accurate segmentation network using the BRATS2020 dataset. Our primary goal was to create a U-Net model that retains great performance using advanced preprocessing techniques. We used a 3D unsharp masking approach in conjunction with normalization to improve image quality. We also examined two smoothing strategies, Gaussian smoothing and median filter smoothing, before applying the unsharp mask. Our findings show that combining Gaussian smoothing, unsharp masking, and normalization resulted in segmentation accuracies of 91.19% for the entire tumor, 85.70% for the enhanced tumor, and 81.80% for the tumor core. In comparison, the median filter smoothing strategy achieved accuracies of 90.78% for the entire tumor, 84.56% for the enhanced tumor, and 83.33% for the tumor core, which is comparable to other results in literature. These findings demonstrate the potential of our preprocessing technique to improve the performance of convolutional neural networks in brain tumor segmentation tasks, making an important addition to the area of medical image analysis.