Enhancing Brain Tumor MRI Classification Using Transfer Learning and Diffusion-Based Augmentation

Abstract

Brain tumor classification using magnetic resonance imaging (MRI) is essential for early diagnosis, but its performance is often limited by small and imbalanced datasets. This study uses 196 MRI images from Kaggle for binary classification. To address data scarcity, multiple approaches are evaluated, including traditional augmentation, patch-based learning, diffusion-based augmentation, and transfer learning. A baseline CNN achieves 87% accuracy but suffers from overfitting. Conventional augmentation provides limited improvement, while patch-based and diffusion-based methods negatively affect performance due to label noise and loss of structural details. In contrast, transfer learning with MobileNetV2 significantly improves generalization. The results show that MobileNetV2 achieves the best performance, highlighting that preserving data quality and leveraging transfer learning are more effective than simply increasing dataset size.

Introduction

Medical image classification, particularly brain tumor detection using MRI scans, is highly important for early diagnosis but faces a major challenge due to the limited availability of labeled medical data. This study works with a very small dataset of 196 MRI images, which leads to overfitting when training deep learning models.

A baseline CNN model is first developed, achieving around 87% accuracy but showing clear overfitting. Traditional data augmentation methods (rotation, flipping, scaling) provide only slight improvements because they do not add meaningful new information. To increase data size, patch-based learning is used, expanding the dataset to over 3000 samples, but it introduces noise since many patches do not contain tumor regions, reducing model effectiveness.

Transfer learning using MobileNetV2 proves to be the most effective approach, as it leverages pre-trained features and significantly improves generalization while reducing overfitting. Diffusion-based data augmentation is also tested, but it produces unrealistic MRI images that blur important tumor structures and ultimately reduce performance.

The literature review highlights that medical AI models depend heavily on large, high-quality, and well-labeled datasets, which are difficult to obtain due to privacy, cost, and expert limitations. Various methods like GANs, diffusion models, advanced augmentation techniques, and even large language models have been explored to generate or enhance data, but challenges remain regarding image quality, clinical realism, computational cost, and model reliability.

Conclusion

Just adding more training samples wasn’t the answer. When new data doesn’t capture the real features of brain tumors, it actually makes things worse. Of all the approaches tested, transfer learning was hands-down the best way to deal with not having much data. The bigger lesson here is that you can’t ignore quality or clinical meaning when building AI for healthcare. If your method doesn’t maintain anatomical or diagnostic information, it won’t be trustworthy.

References

[1] T.-T. Yang, C.-W. Ma, C.-H. Lee, S.-X. Qiu, and C.-C. E. Lan, \"AI grading of lateral canthal lines: Novel models for unseen synthetic image generation and data augmentation,\" Clinical, Cosmetic and Investigational Dermatology, 2024. [2] S. Islam, M. T. Aziz, H. R. Nabil, J. Rahman, M. F. Mridha, M. M. Kabir, N. Asai, and J. Shin, \"Generative adversarial networks (GANs) in medical imaging: Advancements, applications, and challenges,\" IEEE Access, 2024. [3] S. Chatterjee and Y.-C. Byun, \"Leveraging generative adversarial networks for data augmentation to improve fault detection in wind turbines with imbalanced data,\" Results in Engineering, 2024. [4] A. Kumar, P. P. Kumar, and O. Prakash, \"Exploring generative adversarial networks for synthetic data generation in healthcare applications,\" NeuroQuantology, vol. 19, no. 10, pp. 523–532, 2021. [5] B. K. Kalejahi, S. Meshgini, and S. Danishvar, \"Segmentation of brain tumor using a 3D generative adversarial network,\" Diagnostics, vol. 13, 2023. [6] K. Faryna, J. van der Laak, and G. Litjens, \"Automatic data augmentation to improve generalization of deep learning in H&E stained histopathology,\" Computers in Biology and Medicine, vol. 170, 2024. [7] D. Sun, F. Dornaika, and J. Charafeddine, \"LCAMix: Local-and-contour aware grid mixing based data augmentation for medical image segmentation,\" Information Fusion, vol. 110, 2024. [8] D. Sun, F. Dornaika, and N. Barrena, \"HSMix: Hard and soft mixing data augmentation for medical image segmentation,\" Information Fusion, vol. 115, 2025. [9] L. Juwara, A. El-Hussuna, and K. El Emam, \"An evaluation of synthetic data augmentation for mitigating covariate bias in health data,\" Patterns, vol. 5, 2024. [10] B. Ding, C. Qin, R. Zhao, T. Luo, X. Li, G. Chen, W. Xia, J. Hu, A. T. Luu, and S. Joty, \"Data augmentation using LLMs: Data perspectives, learning paradigms and challenges,\" Findings of ACL, 2024. [11] K. G. A. Villanueva and P. Kumar, \"XAI-driven skin disease classification: Leveraging GANs to augment ResNet-50 performance,\" arXiv preprint, 2025. [12] \"Data augmentation strategies for GAN-based medical image analysis: An empirical review,\" 2023. [13] \"Diffusion-based generative model for medical image augmentation,\" 2024.

Copyright

Copyright © 2026 Rohan Shinde, Janhavi Manjarekar, Vishal Akale, Pramod Jadhavo , Pritish Bisne. 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.