Robust Brain Tumor Classification using Attention-Enhanced Swin Transformers

Official implementation of the paper: "Robustness of Attention-Enhanced Swin Transformers Against Data Leakage in Brain Tumor Classification".

Overview

State-of-the-art brain tumor classification models often suffer from Data Leakage due to random splitting, leading to inflated accuracies (99%+) that collapse on unseen patients.

We propose a Patient-Level Split benchmark and a novel Attention-Enhanced Swin Transformer that integrates CBAM to strictly learn pathological features, achieving 96.82% accuracy on the rigorous "Hard Mode" split.

---

Dataset Link : Brain Tumor Dataset on Figshare

---

Key Features

• Strict Patient-Level Split: Zero overlap of patients between Train/Test sets.
• Attention Mechanism: Integrated CBAM (Channel + Spatial Attention) into Swin-Tiny.
• SOTA Performance: Outperforms EfficientNet-B0 and ViT-Small on unseen patients.
• Explainability: Grad-CAM visualizations confirming artifact-free learning.

---

Results (Patient-Level Split)

Model	Parameters	Accuracy (Unseen Patients)	Robustness Gap
ResNet-50	25.6M	95.97%	2.29%
EfficientNet-B0	5.3M	91.95%	3.71%
ViT-Small	22.0M	94.28%	3.12%
Swin-Tiny	28.3M	96.19%	1.64%
Att-Swin (Ours)	28.3M	96.82%	1.23%

---

Qualitative Analysis

Our model ignores background noise (skull, eyes) and focuses strictly on the tumor mass, as confirmed by Grad-CAM.

---

Citations:

If you use this code, please cite our paper:

@inproceedings{chhetri2026robust, title={Robustness of Attention-Enhanced Swin Transformers Against Data Leakage in Brain Tumor Classification}, author={Chhetri, Latchan and Ghosal, Palash}, booktitle={IEEE Conference Name (GCON)}, year={2026} }

Contacts:

For questions, please contact Latchan Chhetri.

Installation

git clone [https://github.com/Latchan-Ch/Robust-Swin-Brain-Tumor.git](https://github.com/Latchan-Ch/Robust-Swin-Brain-Tumor.git)
cd Robust-Swin-Brain-Tumor