Multi-Class Brain Tumor Classification from MRI Using ResNet50V2 Transfer Learning with Two-Stage Augmentation, Grad-CAM Explainability, and Cross-Dataset Validation

Abstract

Brain tumor classification from magnetic resonance imaging (MRI) remains a challenging clinicaltaskduetointer-classvisualsimilarityandsignificantvari-ation across imaging protocols and acquisition equipment.Accurate and timely classification of glioma, meningioma, pituitary tumor, and normal brain tissue is critical for treatment planning and patient outcomes.This work proposes a two-phase transfer learning framework based on ResNet50V2 pretrained on ImageNetfor four-class brain MRI classification.A two-stage augmentation pipeline combin-ingofflinedegradationaugmentationGaussiannoise(SNR=18.39dB),Gaussianblur(SSIM?0.93),andresolutiondegradation(SSIM?0.85)—withonlinegeometric augmentation was employed to improve cross-scanner robustness. Class imbalance was addressed through balanced class weighting.Grad-CAM visualisa-tions were generated to provide anatomically interpretable explanations of model decisions.Evaluated on a held-out test set of 310 images from the Daneshmand dataset, the proposed model achieved 89.68% accuracy, Cohen’s ?=0.8595, and macro-AUCof0.9820. ExternalvalidationontheindependentCE-MRIdataset (n=394)yielded84.52%accuracyandmacro-AUCof0.9729, withMcNemar’s test confirming no statistically significant difference in error patterns (p=0.1882), demonstrating consistent cross-dataset generalisation. Grad-CAM activations con-firmed anatomically consistent focus across all four classes, supporting clinical in-terpretability.The complete implementation is publicly available at the project repository.

Introduction

This paper presents a deep learning framework for brain tumor classification from MRI images, aiming to improve accuracy, generalization, and clinical interpretability across four classes: glioma, meningioma, pituitary tumor, and normal brain.

Key motivation

Brain tumors are life-threatening and require early and accurate classification, but manual MRI interpretation is:

• Time-consuming
• Subject to inter-observer variation
• Challenged by large imaging volumes

A major difficulty is that tumor types (especially glioma vs meningioma) often have visually similar MRI patterns, and datasets are typically imbalanced and limited.

Proposed method

The study introduces a transfer learning–based CNN approach using ResNet50V2, combined with a carefully designed training pipeline:

• A two-stage augmentation system (noise, blur, resolution degradation, and geometric transforms) validated using SSIM, PSNR, and SNR to ensure clinical realism.
• A fine-tuned ResNet50V2 model with transfer learning:
  • First trained with frozen layers
  • Then fine-tuned on deeper layers
• Class imbalance handling using weighted loss functions
• Grad-CAM explainability to visualize which brain regions influenced predictions

Dataset and validation

• Primary dataset: 3,096 MRI images
• External validation: 394 images (CE-MRI dataset)
• Strict train/validation/test split with no data leakage

Results

• Internal test accuracy: 89.68%
• Macro AUC: 0.9820
• External dataset accuracy: 84.52% (without retraining)
• Strong generalization (minimal AUC drop of ~0.009)

Key findings

• Best performance achieved for normal and pituitary classes
• Main errors occur between glioma and meningioma, due to overlapping imaging features
• Grad-CAM shows the model focuses on clinically meaningful brain regions
• Statistical tests confirm stable and consistent generalization across datasets

Comparison with prior work

The model performs competitively with or better than earlier CNN, SVM, and transfer learning approaches, with a key advantage being:

• External dataset validation
• Augmentation quality validation
• Explainable AI (Grad-CAM)

Limitations

• Uses 2D MRI slices (no 3D context)
• Relies on only T1-weighted scans (limited anatomical detail)
• Training data is from a single source
• No prospective real-world clinical deployment yet

Conclusion

AtwophaseResNet50V2transferlearningframeworkwaspresentedforfourclassbrainMRItumorclassification,achieving89.68%accuracy, ?=0.8595, andmacro-AUC0.9820 on a strictly held-out internal test set. External validation on the independent CE-MRI benchmarkachieved84.52%accuracyandmacroAUC0.9729,withMcNemar’stestconfirmingnostatisticallysignificantdifferenceinerrorpatterns(p=0.1882), demonstrating consistent cross-dataset generalisation.Five principal contributions were demonstrated: quantitatively validated two-stage augmentation with mathematical formulation, two-phase transfer learning with progressive fine-tuning, balanced class weighting, anatomi-cally consistent Grad-CAM explainability with formal derivation, and statistically vali-dated external cross-dataset generalisation.

References

[1] D. N. Louis et al., “The 2021 WHO Classification of Tumors of the Central Nervous System:asummary,”Neuro-Oncology,vol.23,no.8,pp.1231–1251,2021. [2] Q.T.Ostrometal.,“CBTRUSStatisticalReport:PrimaryBrainandOtherCen-tral Nervous System Tumors Diagnosed in the United States in 2014–2018,” Neuro-Oncology, vol. 23, no. S3, pp. iii1–iii105, 2021. [3] D. N. Louis et al., “The 2016 World Health Organization Classification of Tumors ofthe Central Nervous System,” Acta Neuropathologica, vol. 131, no. 6, pp. 803–820, 2016. [4] A. Salehi, A. Ahmadi, and I. Heidari, “A brief survey on deep learning based brain tumor detection,” arXiv preprint arXiv:2007.01169, 2020. [5] C. Shorten and T. M. Khoshgoftaar,“A survey on image data augmentation fordeep learning,” Journal of Big Data, vol. 6, no. 1, pp. 1–48, 2019. [6] J.Chengetal.,“Enhancedperformanceofbraintumorclassificationviatumor regionaugmentationandpartition,”PLOSONE,vol.10,no.10,p.e0140381,2015. [7] N.Abiwinanda,M.Hanif,S.T.Hesaputra,A.Handayani,andT.R.Mengko,“Brain tumorclassificationusingconvolutionalneuralnetwork,”inWorldCongressonMed-ical Physics and Biomedical Engineering, Singapore:Springer, 2019, pp. 183–189. [8] A. Pashaei, H. Sajedi, and N. Jazayeri, “Brain tumor classification via convolutional neural network and extreme learning machines,” in Proc. 8th Int. Conf. Computer and Knowledge Engineering (ICCKE), 2018, pp. 314–319. [9] S.DeepakandP.M.Ameer,“BraintumorclassificationusingdeepCNNfeatures via transfer learning,”ComputersinBiologyandMedicine,vol. 111,p. 103345,2019. [10] M. A. Khan et al., “Brain tumor detection and classification using MRI: a cross-sectional study,” Computational and Mathematical Methods in Medicine, vol. 2020, [11] p.6276735,2020. [12] A. Rehman, S. Naz, M. I. Razzak, F. Akram, and M. Imran, “A deep learning-based framework for automatic brain tumor classification using transfer learning,” Circuits, Systems,andSignalProcessing,vol.39,no.2,pp.757–775,2020. [13] H. H. Sultan, N. M. Salem, and W. Al-Atabany, “Multi-classification of brain tumor imagesusingdeepneuralnetwork,”IEEEAccess,vol.7,pp.69215–69225,2019. [14] S.Daneshmand,“BrainMRIImagesforBrainTumorDetection,”Kaggle,2020. [Online]. Available: https://www.kaggle.com/datasets/susandaneshmand/mri-images [15] S.Chakrabarty,“BrainTumorClassification(MRI),”Kaggle,2019.[On-line].Available: https://www.kaggle.com/datasets/sartajbhuvaji/brain-tumor-classification-mri [16] K.He, X.Zhang, S.Ren, andJ.Sun, “Identitymappingsindeepresidualnetworks,” in Proc. ECCV, Amsterdam, 2016, pp. 630–645. [17] R.R.Selvarajuetal.,“Grad-CAM:Visualexplanationsfromdeepnetworksvia gradient-based localization,” in Proc. ICCV, Venice, 2017, pp. 618–626. [18] D.P.KingmaandJ.Ba,“Adam:Amethodforstochasticoptimization,”inProc. ICLR, 2015. [19] J. R. Landis and G. G. Koch, “The measurement of observer agreement for categor-ical data,” Biometrics, vol. 33, no. 1, pp. 159–174, 1977. [20] S.Muji,“MRIBrainTumorClassification,”Kaggle,2025.[Online].Available:https://www.kaggle.com/code/shalmamuji/mri-brain-tumor-classification

Copyright

Copyright © 2026 Shalma Muji M, Nithin T, Lavanya D, Abhiram VS, S Sreeja, A Kousalya. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.