Explainable Vision Transformer–CNN Fusion Network for Brain Tumor Detection and Classification using MRI with Grad-CAM Visualization

View at publisher Summarise with VIRAT
Authors: Vishal, Shiva Mehta, Vinay Kukreja
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Magnetic resonance imaging (MRI) is an essential part of the early diagnosis and management of treatment through accurate and explainable brain-tumor classification. The paper is a hybrid Vision Transformer-Convolutional Neural Network (ViT-CNN) architecture, which features both the local feature extraction power of CNNs and the global attention of Vision Transformers, with the interpretation of Grad-CAM visualization. The model was trained and tested on the Br35H and Figshare MRI datasets which had four types of tumors- glioma, meningioma, pituitary and no tumor. The experimental data show that the proposed ViT-CNN achieved an excellent accuracy of 98.72, precision of 98.23, recall of 98.41, F1-score of 98.31, and the mean AUC of 0.991, which is superior to the baseline models such as ResNet-50 (96.1 percent), DenseNet-121 (96.5 percent), and pure ViT (97.1 percent). The model was also stable and had a standard deviation average of less than 0.25 per cent across five folds and a Cohen k of 0.982, which is almost near perfect in terms of consistency in predicting. ROC analysis showed per-class AUCs of 0.992 (glioma), 0.988 (meningioma), 0.994 (pituitary), and 0.990 (no tumor). The smooth convergence of the training and validation curves were confirmed with a small generalization gap (< 0.02). Grad-CAM images were able to localize tumor regions based on model attention to improve clinical interpretability. The proposed ViT-CNN has 39.8 M parameters, 6.7 GFLOPs, and inference time of 8.3 ms/image, which is an accurate, efficient, and explainable diagnostic framework, thus being much more useful in real-time and reliable medical image analysis.

Fields of Study & Subjects

Brain Tumor Detection and Classification AI in cancer detection Glioma Diagnosis and Treatment

Publication details

Journal / source
International Conference on Image Processing and Capsule Networks (ICIPCN)
Pages
427–432
Publication date
27 January 2026
Language
EN