Attention-Enhanced Swin Transformers for Robust Brain Tumor Classification Under Patient-Level Data Splitting

Abstract

Contemporary brain tumor classification systems report accuracies exceeding 98 percent, yet such metrics are artificially inflated by dataset partitioning flaws. Conventional image-level splitting allows identical patient scans in both training and testing sets, enabling networks to memorize patient-specific features rather than tumor patterns. We enforce patient-level separation where each patient remains in a single partition. Evaluation of five architectures reveals degradations reaching 3.71 percent under patient-level splitting, validating widespread data leakage. We propose an Attention-Enhanced Swin Transformer integrating hierarchical windowed attention with Convolutional Block Attention Modules. Our architecture achieves 96.82 percent accuracy with 1.23 percent degradation—the smallest gap among methods. Gradient-weighted activation mapping confirms attention on tumor regions rather than anatomical artifacts, establishing reliable feature extraction for trustworthy medical AI.

Key Methodologies & Contributions

  • Algorithmic Patient-Level Partitioning: Enforced strict patient-level separation where each patient’s MRI scans remain in a single partition, eliminating data leakage that artificially inflates accuracy by 5-30 percentage points in conventional image-level splitting.
  • Attention-Enhanced Architecture: Developed an Attention-Enhanced Swin Transformer that integrates hierarchical windowed attention with Convolutional Block Attention Modules (CBAM) to refine features through sequential channel and spatial attention.
  • Anatomical Artifact Suppression: Utilized CBAM to explicitly prioritize tumor-discriminative features while suppressing patient-specific anatomical variations, such as skull boundaries and ventricular configurations.
  • Robust Generalization: Achieved 96.82 percent classification accuracy under rigorous patient-level splitting, demonstrating a minimal performance degradation of 1.23 percent compared to image-level evaluation.


Code & Resources

Status: Accepted at IEEE GCON, 2026. Authors: L. Chhetri, A. Datta, P. Ghosal