Hybrid encoder-decoder to preserve long-range context in brain tumor segmentation with EfficientNetB7 and a convolutional vision transformer mechanism

Accurate segmentation of brain tumors in MRI is critically important for clinical diagnosis, particularly challenging for Low-Grade Gliomas (LGGs) in FLAIR sequences where precise delineation directly impacts treatment planning and patient monitoring. Conventional U-Net architectures face limitations in this domain, exhibiting structural degradation from repetitive down-sampling and inefficient feature propagation between encoder and decoder components. Furthermore, most existing approaches rely on manual hyperparameter optimization, which is both computationally intensive and prone to suboptimal convergence. To address these challenges, this study introduces a novel hybrid encoder-decoder framework that integrates an EfficientNetB7 backbone with a single-stage Convolutional Vision Transformer (CvT) and a TransUNet style decoder, systematically optimized through Bayesian hyperparameter tuning. The architecture leverages EfficientNetB7 for multi-scale feature extraction while the CvT module implements convolutional token embedding and multi-head self-attention to model long-range dependencies without patch-based transformations, maintaining structural coherence for optimized feature fusion. The decoding pathway employs bilinear upsampling enhanced with residual convolutional blocks for precise boundary refinement. Extensive evaluation on the Kaggle Low-Grade Glioma (LGG) and Brain Tumor Segmentation datasets demonstrates superior performance, achieving Dice coefficients of 0.9290 and 0.9262 respectively. These results confirm that our architecture effectively preserves fine structural details while capturing global context, yielding robust and accurate tumor segmentation, demonstrating strong potential for future integration into clinical decision support systems, pending prospective clinical validation. The integration of Bayesian optimization further ensures reproducible and optimal model configuration, enhancing the framework’s reliability for medical imaging applications.