ADA-XAI: Adaptive Faithfulness-Driven Explainability for Brain Tumor MRI Classification

Adaptive Explainable AI Framework for Brain Tumor Diagnosis using Hybrid EVA-02 Transformers and Deformable CNNs

---

Overview

ADA-XAI is a hybrid Explainable AI framework designed for brain tumor classification from MRI scans using a combination of:

• EVA-02 Vision Transformers
• Deformable CNNs (DCNN)
• Cross Attention Fusion
• Adaptive Explainability Mechanism

Unlike conventional CNN-only or Transformer-only systems, ADA-XAI dynamically combines:

• Global semantic understanding from EVA-02
• Local irregular tumor morphology extraction from DCNN

The framework also introduces an Adaptive XAI mechanism that intelligently selects between:

• Grad-CAM
• CAD-GMAR (Cross Attention Gradient-Driven Multi-Head Attention Rollout)

based on real-time faithfulness evaluation.

The proposed model achieved 99% accuracy on the Kaggle Brain Tumor MRI dataset.

---

Published Research Paper

Ada-XAI: Adaptive Faithfulness-Driven Explainability for Hybrid Eva-02 and Deformable CNNs in Brain Tumor Diagnosis From MRI Images

Published at:

2026 International Conference on Innovative Trends in Information Technology (ICITIIT)

DOI: https://doi.org/10.1109/ICITIIT68860.2026.11499683

---

Architecture

The framework consists of:

1. EVA-02 Global Context Branch

• Vision Transformer pretrained using Masked Image Modeling

• Captures:

  • Long-range dependencies
  • Global anatomical structure
  • Semantic context

2. Deformable CNN Local Branch

• Uses deformable convolutions

• Learns:

  • Irregular tumor boundaries
  • Local texture patterns
  • Shape-adaptive receptive fields

3. Cross Attention Fusion

The model fuses CNN and Transformer features using a Query-Key-Value attention mechanism:

• Query → CNN features
• Key & Value → EVA-02 features

This enables:

• Local feature enhancement using global context
• Better localization
• Robust semantic understanding

4. Adaptive Explainability (ADA-XAI)

The system dynamically chooses the best XAI method using a faithfulness-driven selection mechanism.

Selected Methods

• Grad-CAM
• CAD-GMAR

Selection Criteria

• Perturbation AUC
• Confidence drop analysis
• Dominance evaluation between Transformer and CNN branches

---

Features

• Hybrid Transformer + CNN architecture
• Adaptive explainability
• Faithfulness-driven XAI selection
• MRI tumor classification
• Robust against noisy MRI inputs
• Transfer learning pipeline
• Research-oriented modular design
• High interpretability for medical AI

---

Results

Model	Accuracy
KNN	82%
SVM	88%
CNN	96.03%
ViT	98.86%
ADA-XAI (Proposed)	99%

The framework demonstrated:

• Improved explainability fidelity
• Better localization performance
• Higher robustness to MRI noise perturbations

---

Adaptive XAI Workflow

The framework generates:

• Mcnn using Grad-CAM
• Mvit using CAD-GMAR

A faithfulness metric evaluates both explanations in real time and selects the optimal explanation map.

---

Dataset

The framework uses:

Kaggle Brain Tumor MRI Dataset

• 7023 MRI images

• 4 classes:

  • Glioma
  • Meningioma
  • Pituitary
  • No Tumor

Figshare Brain Tumor Dataset

• 3064 MRI samples
• Transfer learning pretraining stage

---

Explainability Output

The system automatically selects:

• Grad-CAM when CNN dominates
• CAD-GMAR when Transformer dominates

This enables:

• More faithful explanations
• Improved stability
• Better clinical trustworthiness

---

Research Contributions

Proposed Contributions

• Adaptive faithfulness-driven XAI
• Hybrid EVA-02 + DCNN fusion
• Cross-attention-based feature integration
• Transformer-aware explainability
• Robust MRI classification pipeline

---

Citation

@INPROCEEDINGS{11499683,
  author={Sreehari R and Abhinav R and Kalidas V.S and Neeraj Sukumaran and Rajeev Rajan and Chinchu M S},
  booktitle={2026 International Conference on Innovative Trends in Information Technology (ICITIIT)},
  title={Ada-XAI: Adaptive Faithfulness-Driven Explainability for Hybrid Eva-02 and Deformable CNNs in Brain Tumor Diagnosis From MRI Images},
  year={2026},
  doi={10.1109/ICITIIT68860.2026.11499683}
}

---

Future Work

• Clinical-grade evaluation
• Multi-modal MRI support
• Real-time deployment
• Federated medical AI
• 3D MRI explainability
• Explainable segmentation framework

---

Author

• Sreehari R

Co-Author

• Abhinav R
• Kalidas V.S
• Neeraj Sukumaran
• Rajeev Rajan
• Chinchu M S

---

License

This project is licensed under the MIT License.

---

Acknowledgements

• ICITIIT 2026
• Kaggle Brain Tumor Dataset
• Figshare Brain Tumor Dataset

---

Repository

ADA-XAI-MRI Repository