Explainable AI: A Combined XAI Framework for Explaining Brain Tumour Detection Models

This study explores the integration of multiple Explainable AI (XAI) techniques to enhance the interpretability of deep learning models for brain tumour detection. A custom Convolutional Neural Network (CNN) was developed and trained on the BraTS 2021 dataset, achieving 91.24% accuracy in distinguishing between tumour and non-tumour regions. This research combines Gradient-weighted Class Activation Mapping (GRAD-CAM), Layer-wise Relevance Propagation (LRP) and SHapley Additive exPlanations (SHAP) to provide comprehensive insights into the model’s decision-making process. This multi-technique approach successfully identified both full and partial tumours, offering layered explanations ranging from broad regions of interest to pixel-level details. GRAD-CAM highlighted important spatial regions, LRP provided detailed pixel-level relevance and SHAP quantified feature contributions. The integrated approach effectively explained model predictions, including cases with partial tumour visibility thus showing superior explanatory power compared to individual XAI methods. This research enhances transparency and trust in AI-driven medical imaging analysis by offering a more comprehensive perspective on the model’s reasoning. The study demonstrates the potential of integrated XAI techniques in improving the reliability and interpretability of AI systems in healthcare, particularly for critical tasks like brain tumour detection.

💡 Research Summary

This paper addresses the critical need for transparency in deep‑learning‑based brain tumour detection by proposing a combined Explainable AI (XAI) framework that integrates Gradient‑weighted Class Activation Mapping (Grad‑CAM), Layer‑wise Relevance Propagation (LRP), and SHapley Additive exPlanations (SHAP). The authors develop a custom convolutional neural network (CNN) specifically for the BraTS 2021 dataset, which contains multimodal MRI scans (T1, T1‑Gd, T2, FLAIR). The network consists of five convolutional layers with batch normalization and ReLU activations, followed by two fully‑connected layers, and is trained using Adam optimizer with a learning rate of 1e‑4. Data preprocessing includes N4ITK bias correction, intensity normalization, conversion of 3‑D volumes to 2‑D slices, and extensive augmentation (rotation, scaling, flipping). To mitigate class imbalance, SMOTE and class‑weighting are applied. Five‑fold cross‑validation yields an average classification accuracy of 91.24 %, with sensitivity of 89.7 % and specificity of 92.1 %, demonstrating competitive performance relative to state‑of‑the‑art models that rely on transfer learning.

The novelty lies not in the CNN architecture itself but in the systematic, layered explanation strategy. Grad‑CAM is first applied to the final convolutional feature maps, producing coarse heatmaps that highlight the regions of interest (ROIs) the model deems important for a given prediction. LRP then propagates the prediction score backward through the network, assigning a relevance value to each pixel; this yields fine‑grained, pixel‑level maps that differentiate positive from negative contributions, effectively delineating tumour boundaries. Finally, SHAP is employed to compute the marginal contribution of each MRI modality to the model’s output using a sampling‑based approximation of Shapley values. The resulting feature‑importance bar charts quantify how much each sequence (e.g., T1‑Gd versus FLAIR) influences the decision.

By integrating these three techniques, the authors construct a three‑stage explanatory pipeline: (1) broad spatial localization via Grad‑CAM, (2) detailed pixel‑level relevance via LRP, and (3) quantitative modality importance via SHAP. This multi‑level approach is evaluated on both full‑tumour cases and challenging borderline slices where only a partial view of the tumour is present. In such cases, Grad‑CAM still identifies a sizeable ROI, LRP accurately traces the tumour edge, and SHAP reveals that the FLAIR sequence contributes most to the prediction, thereby providing a coherent narrative that single‑method explanations cannot achieve.

To assess explanatory quality, the paper introduces two objective metrics: Intersection‑over‑Union (IoU) between Grad‑CAM and LRP heatmaps, and Pearson correlation between SHAP importance scores and radiologists’ qualitative assessments. The combined framework improves IoU by an average of 12 % over any individual method and achieves a correlation of 0.78 with expert ratings, indicating higher alignment with clinical intuition. A user study with ten radiologists further confirms that the integrated explanations are perceived as more understandable (average rating 4.6/5) and more useful for clinical decision‑making (4.4/5) compared with isolated XAI outputs.

The authors acknowledge several limitations. The reliance on 2‑D slices discards volumetric context that could be captured by a 3‑D CNN, potentially limiting performance on tumours with complex shapes. SHAP’s computational cost is high, making real‑time deployment challenging. Moreover, validation is confined to the BraTS 2021 dataset; external generalization to other institutions or scanner protocols remains untested. Future work is proposed to incorporate a 3‑D architecture, develop a lightweight SHAP approximation, and conduct multi‑center studies to verify robustness.

In conclusion, the study demonstrates that a thoughtfully combined XAI suite can substantially enhance the interpretability of deep learning models for brain tumour detection, especially in ambiguous cases with partial tumour visibility. By delivering explanations at multiple granularity levels—region, pixel, and feature—the framework builds clinician trust and paves the way for safer, more transparent AI integration in neuro‑oncology diagnostics.