Related Experiment Videos
Deep Learning for Brain Tumour Analysis: A Systematic Review of CNN-Transformer Hybrids in Multimodal Imaging
Solomon Buabeng Antwi1,2, Peter Appiahene3, Ben Beklisi Kwame Ayawli2
1Department of Computer Science and Informatics, University of Energy and Natural Resources, Sunyani, Ghana, uenr.edu.gh.
International Journal of Biomedical Imaging
|June 18, 2026
Summary
Hybrid CNN-Transformer models show promise for brain tumour detection, offering higher accuracy than CNN- or Transformer-only approaches. Further research is needed to address computational efficiency and generalizability for clinical use.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Brain Tumour Analysis
- Computational Neuroscience
Background:
- Brain tumour detection requires both local spatial features (CNNs) and global context (Transformers).
- Optimal deep learning architectures for clinical brain tumour detection remain unclear.
- This review examines hybrid CNN-Transformer models, GANs, and multimodal fusion for brain tumour analysis.
Purpose of the Study:
- To systematically review and meta-analyze hybrid CNN-Transformer architectures for brain tumour detection.
- To evaluate diagnostic accuracy, computational efficiency, and the role of GANs and multimodal fusion.
- To identify optimal integration strategies and future research directions.
Main Methods:
- Systematic search of four databases (Jan 2021-May 2025) yielding 94 studies.
- Quality assessment using QUADAS-2 and ROBINS-I frameworks.
- Random-effects meta-analysis of diagnostic accuracy and computational efficiency (GFLOPs).
Main Results:
- Pooled diagnostic accuracy was 93.5%, with hybrid models showing a mean accuracy of 94.6%.
- Parallel hybrid architectures offered the best accuracy-efficiency balance (94.3% accuracy, 2.8 GFLOPs).
- GANs improved rare tumour detection (7%-10%); multimodal fusion achieved 94.2% accuracy.
Conclusions:
- Hybrid CNN-Transformer architectures may enhance diagnostic accuracy for brain tumours.
- Parallel integration strategies show the most favorable accuracy-efficiency balance.
- Challenges in computational efficiency, noise robustness, and generalization require further investigation.