Towards clinically reliable AI: comparative evaluation of CNN, transformer, and hybrid architectures for meningioma detection and segmentation.
Authors
Affiliations (5)
Affiliations (5)
- Gamma Knife Unit, Department of Neurosurgery, Faculty of Medicine, Gazi University, Ankara, 06100, Turkey.
- Institute of Artificial Intelligence, Ankara University, Ankara, 06100, Turkey.
- Department of Software Engineering, Faculty of Engineering, Ankara University, Ankara, 06830, Turkey.
- Department of Biology, Faculty of Science, Ankara University, Ankara, 06100, Turkey. [email protected].
- Department of Biology, Faculty of Engineering and Natural Sciences, Osmaniye Korkut Ata University, 80000, Osmaniye, Turkey. [email protected].
Abstract
Accurate and clinically reliable meningioma detection from brain MRI remains a critical challenge, requiring not only high predictive performance but also robust interpretability. In this study, we systematically evaluated convolutional neural networks (CNNs), Transformer-based models, and hybrid CNN-Transformer architectures for meningioma classification using a unified experimental protocol. A dataset of over 16.000 augmented T1-weighted MRI slices from 100 patients was constructed under ethical approval, with anonymized imaging and patient consent. Classification performance was assessed using accuracy, precision, recall, F1-score, specificity, ROC-AUC, calibration curves, and Wilcoxon signed-rank tests. Transformer models achieved near-perfect performance, with CaiT reaching 99.04% accuracy (AUC = 1.00) and DeiT-Tiny reaching 99.00% accuracy (AUC = 0.99). CNN-based GhostNet achieved 94% accuracy, while HardCoreNAS-C showed lower performance (87%). Hybrid models significantly enhanced weaker CNN backbones, with HardCoreNAS-C + DeiT-Tiny improving accuracy from 87 to 99%. For segmentation, U-Net demonstrated superior performance (Dice = 0.96, IoU = 0.91), outperforming DeepLabV3 + (Dice = 0.80, IoU = 0.62). Explainable AI analysis using LIME revealed that high accuracy alone does not ensure clinical interpretability, as Transformer-based models often highlighted broad brain regions. Overall, this study provides a comprehensive benchmark demonstrating the superiority of Transformer models in classification and U-Net in segmentation, while emphasizing the necessity of interpretability-aware evaluation for clinically deployable AI systems.