TumorXAI: Self-Supervised Deep Learning Framework for Explainable Brain MRI Tumor Classification

Technical Summary: TumorXAI Framework for Brain MRI Classification

Problem Statement
The automatic classification of brain tumors using Magnetic Resonance Imaging (MRI) is critical for early diagnosis and treatment. However, the field faces significant challenges due to the high heterogeneity of tumors and the scarcity of large-scale, annotated medical datasets. Traditional supervised deep learning approaches rely heavily on labeled data, which is expensive and time-consuming to obtain, particularly for rare tumor types that lead to class imbalance. Consequently, there is a need for methods that can leverage abundant unlabeled data to learn robust feature representations without explicit supervision.

Methodology
The authors propose TumorXAI, a self-supervised learning (SSL) framework designed for multi-class brain tumor classification. The methodology follows a structured pipeline:

1. Dataset: The study utilizes a publicly available dataset of 4,448 MRI scans comprising 17 distinct tumor classes (including Glioma, Schwannoma, Meningioma, Neurocitoma, Other Lesions, and Normal cases). The dataset exhibits inherent class imbalance.
2. Preprocessing and Augmentation: To enhance model robustness, the authors applied strong data augmentation techniques, including random resizing, cropping, horizontal flipping, rotation (±20°), zooming (±10%), brightness adjustment, and Gaussian blur. The dataset was augmented to balance classes, resulting in a training set of 8,500 images, a validation set of 1,062, and a test set of 1,063.
3. Model Architecture: The framework employs a ResNet-50 backbone. The authors evaluated four prominent SSL frameworks for pretraining: SimCLR, BYOL, DINO, and MoCo v3.
   • SimCLR: Utilizes a contrastive learning approach with a projection head (Linear → ReLU → Linear) and NT-Xent loss to maximize agreement between augmented views of the same image.
   • BYOL, DINO, MoCo v3: These were implemented with specific hyperparameter configurations (learning rates, temperature settings, and weight decays) tailored to each algorithm.
4. Training Pipeline: The workflow involves SSL pretraining (80 epochs), followed by linear evaluation (50 epochs) and fine-tuning (50 epochs). The models were optimized using the AdamW optimizer with cross-entropy loss and early stopping.
5. Explainability (XAI): To ensure clinical interpretability, the study integrates Explainable AI techniques, specifically Grad-CAM, Grad-CAM++, and Eigen-CAM, to visualize the regions of the MRI images that influence the model's decisions.

Key Contributions

• Comparative Analysis of SSL in Medical Imaging: The paper provides an exhaustive comparative analysis of four state-of-the-art SSL frameworks (SimCLR, BYOL, DINO, MoCo v3) specifically for the complex task of multi-class brain tumor classification (17 classes).
• Performance with Limited Labels: The study demonstrates that SSL-pretrained models can outperform traditional supervised baselines in scenarios where labeled data is limited, effectively addressing the data scarcity issue in medical imaging.
• Integration of XAI: The framework combines high-performance classification with visual interpretability tools, bridging the gap between deep learning accuracy and clinical trustworthiness.
• Robust Feature Learning: The research validates that SSL methods can learn semantic and discriminative feature spaces that effectively separate complex tumor subtypes, as evidenced by t-SNE visualizations.

Results
Experimental evaluations on the 17-class dataset yielded the following key findings:

• SimCLR Superiority: Among the four frameworks, SimCLR achieved the highest performance across all metrics.
  • Fine-tuning Performance: SimCLR reached 99.64% accuracy, precision, recall, and F1-score.
  • Testing Performance: SimCLR maintained a testing accuracy of 97.27% with an F1-score of 0.9727.
  • Validation Performance: It achieved 97.65% accuracy on the validation set.
• Comparison with Other SSL Models:
  • MoCo v3 performed strongly with 99.48% accuracy in fine-tuning but slightly trailed SimCLR in testing (94.92%).
  • BYOL achieved 97.66% accuracy in fine-tuning and 96.14% in testing.
  • DINO showed high fine-tuning accuracy (98.68%) but struggled significantly during linear evaluation (45.24%) and testing (95.58%), indicating potential instability in feature transfer for this specific task.
• Visualization: t-SNE visualizations confirmed that SimCLR produced the most well-defined clusters for different tumor subtypes. Confusion matrix analysis revealed that SimCLR had the least noisy error distribution compared to the other models.
• XAI Validation: The Grad-CAM and Eigen-CAM visualizations successfully highlighted clinically relevant tumor regions, confirming that the model's predictions were based on meaningful anatomical features rather than artifacts.

Significance and Claims
The paper claims that TumorXAI demonstrates the scalability and dependability of self-supervised learning for diagnosing brain tumors from unlabeled medical data. By leveraging SSL, the approach alleviates the dependency on large annotated datasets, a major bottleneck in medical AI. The study positions SimCLR as a robust, scalable tool that can assist radiologists in making efficient and accurate clinical decisions.

The authors note that while the results are promising, the study focused on single-sequence MRI data. They modestly acknowledge that future work should explore multi-sequence and multimodal MRI data to further improve representation and generalization across diverse datasets. The paper concludes that SSL, particularly SimCLR, represents a significant step toward trusted, automated brain tumor classification in clinical practice.