Hybrid encoder-decoder with EfficientNetB7 and convolutional vision transformer improves brain tumor segmentation in MRI

Researchers have developed a new hybrid AI model for brain tumor segmentation in MRI scans, achieving high accuracy on two datasets with Dice coefficients of 0.9290 and 0.9262. The system combines an EfficientNetB7 backbone with a convolutional vision transformer to preserve long-range context. The architecture addresses limitations of conventional U-Net models, which suffer from structural degradation and inefficient feature propagation. It uses Bayesian hyperparameter tuning for optimal configuration, and a TransUNet-style decoder with bilinear upsampling and residual blocks for precise boundary refinement. The study focused on Low-Grade Gliomas in FLAIR sequences, where accurate delineation impacts treatment planning. The authors suggest the framework shows potential for clinical decision support systems, though prospective validation is still needed. The article is published under a Creative Commons license.