Scopus Indexed Publications

Paper Details


Title
A CNN-Transformer Ensemble Framework with Explainable AI for Brain Tumor Identification in MRI

Author
Md. Naimul Islam Nuhash, Md. Ali Emam Al Mamun, Md. Jobayer Ahmed, Raja Tariqul Hasan Tusher,

Email

Abstract

Effective treatment planning of a brain tumor requires accurate and timely diagnosis. This study investigates the deep learning and ensemble-based approaches for a multi-class classification problem of brain MRI images into four classes: Glioma, Meningioma, Pituitary, and No Tumor. The original dataset of 1,505 MRI images was extended to 4,515 through controlled upsampling for balancing classes and improving the model generalization. Four state-of-the-art architectures were trained and tested: InceptionV3, Swin-Tiny, MobileNetV3, and Vision Transformer (ViT). As a result, InceptionV3 has proved to have an individual best test accuracy of 98.53%. Explainable AI techniques are applied only to the best performing model in an attempt to explain the prediction behavior of the model, as well as highlight clinically interpretable regions within MRI. With the objective of enhancing the predictive reliability, several ensemble methods are implemented, including Soft Voting and Stacking, both achieving superior test accuracy of 99.26%, outperforming all the single models. From this perspective, the results of the experiment show that ensemble deep learning approaches, in particular, Soft Voting, a very high accuracy with robust results, whereas XAI assessment adds to essential interpretability of results. These strengths make the proposed framework a sound candidate for automated brain tumor diagnosis in clinical settings.


Keywords

Journal or Conference Name
2025 28th International Conference on Computer and Information Technology, ICCIT 2025

Publication Year
2025

Indexing
scopus