Authors:  David Miller1 and Marc Pomplun2

JSA-Vol. 3 (2024),

David Miller1 and Marc Pomplun2

1 Department of Computer Science, University of Massachusetts Dartmouth, Dartmouth, MA 02747, USA.

2 Department of Computer Science, University of Massachusetts Boston, Boston, MA 02115, USA.

* Correspondence: marc@cs.umb.edu

Received: 16 December 2023; Accepted: 19 January 2025; Published: 15 August 2025.

Abstract: Accurate detection and classification of meningioma brain tumors using magnetic resonance imaging remain a challenging task in medical image analysis. Hybrid deep learning frameworks that integrate convolutional neural networks with directional transforms such as the Ridgelet transform have recently demonstrated very high performance on benchmark datasets. However, most existing studies evaluate their models under limited experimental settings and do not address robustness, generalization across institutions, or predictive uncertainty. This paper presents a robust and uncertainty-aware extension of Ridgelet-enhanced hybrid convolutional neural networks for meningioma detection and classification. The proposed framework emphasizes cross-dataset generalization, robustness under domain shift, and uncertainty estimation for clinical decision support. External validation experiments are conducted using heterogeneous MRI datasets acquired from different scanners and institutions. In addition, Bayesian uncertainty modeling is employed to quantify prediction confidence and identify ambiguous cases. Experimental results demonstrate that while domain shift leads to performance degradation, robustness-aware training and uncertainty estimation significantly enhance reliability and clinical safety. This work contributes toward the deployment of trustworthy artificial intelligence systems for brain tumor diagnosis.

Keywords: Meningioma,