HSegFormer: hybrid CNN–transformer with stage attention for brain tumor MRI segmentation

Brain tumor segmentation from contrast-enhanced T1-weighted MRI poses challenges due to heterogeneous appearances and intricate boundaries. This research introduces HSegFormer, a hybrid CNN-Transformer model that merges convolutional feature extraction with transformer-based contextual modeling. The architecture incorporates attention-guided decoding and deep supervision, trained using binary cross-entropy and Dice losses on the BRISC 2025 and Figshare datasets.

The results demonstrate that transformer-based and hybrid models outperform CNN-based models in terms of IoU and Dice scores. Among the methods evaluated, HSegFormer demonstrates the highest segmentation performance while maintaining efficient computational requirements. Ablation studies reveal that stage-wise attention, deep supervision, and the Combo Loss each contribute to the overall performance enhancement.

These findings indicate that integrating convolutional and transformer representations can significantly improve brain tumor segmentation across the assessed datasets. The research, funded by Iran National Science Foundation (INSF) project No 4044422, was conducted by Bahar Niknam, Amirreza Jalili, and Hedieh Sajedi at the Department of Computer Science, School of Mathematics, Statistics and Computer Science, University of Tehran, Tehran, Iran.

During the preparation of this work, the authors utilized ChatGPT-5 to enhance the manuscript's readability and language, with full responsibility for the content ultimately lying with the authors. Springer Nature maintains neutrality regarding jurisdictional claims in published maps and institutional affiliations. This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, permitting non-commercial use and sharing, provided appropriate credit is given to the original authors and source, and a link to the Creative Commons license is included.