Brain tumor segmentation is paramount in medical diagnostics. This study presents a multistage segmentation model consisting of two main steps. First, the fusion of magnetic resonance imaging (MRI) modalities creates new and more effective tumor imaging modalities. Second, the semantic segmentation of the original and fused modalities, utilizing various modified architectures of the U‐Net model. In the first step, a residual network with multi‐scale backbone architecture (Res2Net) and guided filter are employed for pixel‐by‐pixel image fusion tasks without requiring any training or learning process. This method captures both detailed and base elements from the multimodal images to produce better and more informative fused images that significantly enhance the segmentation process. Many fusion scenarios were performed and analyzed, revealing that the best fusion results are attained when combining T2‐weighted (T2) with fluid‐attenuated inversion recovery (FLAIR) and T1‐weighted contrast‐enhanced (T1CE) with FLAIR modalities. In the second step, several models, including the U‐Net and its many modifications (adding attention layers, residual connections, and depthwise separable connections), are trained using both the original and fused modalities. Further, a “Model Selection‐based” fusion of these individual models is also considered for more enhancement. In the preprocessing step, the images are resized by cropping them to decrease the pixel count and minimize background interference. Experiments utilizing the brain tumor segmentation (BraTS) 2020 dataset were performed to verify the efficiency and accuracy of the proposed methodology. The “Model Selection‐based” fusion model achieved an average Dice score of 88.4%, an individual score of 91.1% for the whole tumor (WT) class, an average sensitivity score of 86.26%, and a specificity score of 91.7%. These results prove the robustness and high performance of the proposed methodology compared to other state‐of‐the‐art methods.
