Complementary cross-gated fusion framework for brain tumor segmentation using MR images (CoGFu Net)

Segmentation of brain tumors using magnetic resonance (MR) images is essential for early diagnosis and effective treatment planning, particularly in areas where tumor subregions exhibit diverse intensity patterns and overlapping boundaries between tumor and normal tissues. Existing methods mainly emphasize global and local contextual tumor representation, often overlooking critical features like structural boundary details and inter feature learning.

A novel model that combines complementary feature learning with cross-gated fusion. Complementary learning focuses on both contextual and structural information through two parallel processes: contextual semantic learning and fine edge learning. These learnings extract global semantic context and local edge-aware details. A cross-gated fusion mechanism is introduced in the skip connections to effectively combine these complementary features, adaptively balancing semantic richness with structural precision. A graphical user interface is also developed to facilitate subregion-aware prediction and metric evaluation.

On the BraTS2020 dataset, the proposed model achieved mean dice coefficient (DC) of 0.9489, 0.8581, and 0.8352 in whole tumor (WT), tumor core (TC), and enhanced tumor (ET) regions, respectively, and mean intersection over union (IOU) of 0.8961, 0.8572, and 0.8093 in WT, TC and ET regions. The model achieved DC of 0.9365 and IOU of 0.8893 on the LGG-MRI dataset.

In the comparative analysis, DC improved by 3.3%, 5.7%, and 3.3% on BraTS2020, BraTS2021, and LGG-MRI datasets respectively, demonstrating the robustness and adaptability of the model.

The proposed framework addresses the challenges across multi-class and single-tumor regions and facilitates practical subregion-aware clinical deployment.