Cross-Benchmark Evaluation of an Efficient CNN–Transformer Model for Brain Tumor Subregion Segmentation
Year: 2026
Type: conference-paper
Abstract: Accurate delineation of brain tumor subregions is critical for treatment planning and outcome monitoring, yet manual MRI segmentation is time-consuming and subject to inter-observer variability. This paper presents an efficient CNN-Transformer hybrid framework for automated multiclass brain tumor subregion segmentation using TransUNetB, a lightweight bottleneck Vision Transformer integrated within a U-Net encoder-decoder architecture. The approach is evaluated on two public benchmarks, BraTS 2020 and BraTS 2021, using a unified pipeline that includes denoising, contrast enhancement, slice/patch-based learning, and consistent training settings. TransUNetB achieves strong tumor-only multiclass performance, reaching 97.50% accuracy on BraTS 2020 and 98.81% on BraTS 2021, with competitive F1, MCC, and PR-AUC scores across both datasets. To promote clinical trust, we further incorporate Grad-CAM-based model transparency, generating class-specific heatmaps that highlight regions influencing ED/ET/TC predictions and supporting qualitative verification of segmentation focus. Overall, the proposed framework provides an accurate and interpretable segmentation solution suitable for scalable clinical workflows and future real-time deployment.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204445829