A Unified XAI Framework for Interpreting Deep Learning Models in Brain Tumor Detection

Abstract

Deep neural networks have shown strong potential for analysing MRI scans in brain tumour detection, but their black-boxnaturemakesthereasoningbehindeachpredictionhard to audit — a serious limitation in safety-critical care. This survey reviews recent work that pairs Convolutional Neural Networks (CNNs) with Explainable Artificial Intelligence (XAI) to mitigate this opacity, taking a CNN trained on the BraTS 2021 dataset as the reference setting. We then examine three complementary explanation families — Gradient-weighted Class Activation Mapping(Grad-CAM),Layer-wiseRelevancePropagation(LRP), andSHapleyAdditiveexPlanations(SHAP)—whichrespectively expose salient regions, per-pixel attributions, and feature-level contributions. We argue that fusing these views into a unifiedXAIpipelineyieldslayered,moretrustworthyrationalesthanany single method, and is a practical path toward auditable clinical decision support.

Introduction

This survey examines the integration of Explainable Artificial Intelligence (XAI) with Deep Learning (DL) for brain tumor detection and analysis using MRI images. While Convolutional Neural Networks (CNNs) achieve high accuracy in identifying brain tumors, their decision-making process remains largely opaque, limiting trust and adoption in clinical settings. XAI addresses this issue by making AI predictions interpretable through techniques such as Grad-CAM, Layer-wise Relevance Propagation (LRP), and SHAP, which highlight important image regions, pixels, or features influencing model decisions.

The survey reviews recent research on brain tumor detection, including transfer learning, CNN-based classification, segmentation approaches, hybrid CNN-SVM models, and various XAI methods. Each technique offers unique advantages but also has limitations. Grad-CAM provides region-level visual explanations but lacks fine detail, LRP delivers pixel-level attribution but can be noisy, and SHAP offers feature-level insights at the cost of high computational complexity. Studies consistently show that no single XAI method provides a complete explanation of model behavior.

To overcome these limitations, the survey advocates a unified XAI framework that combines multiple explainability techniques. In such a framework, Grad-CAM identifies suspicious tumor regions, LRP highlights important pixels, and SHAP explains feature contributions simultaneously. This multi-resolution explanation approach improves transparency, clinician confidence, and trustworthiness of AI-assisted diagnosis.

The paper also discusses key challenges in deploying XAI-enhanced medical imaging systems, including the lack of standardized evaluation metrics for explanations, high computational costs, difficulties in real-time deployment, limited clinician-friendly interfaces, robustness concerns, and regulatory requirements. Future research directions include lightweight explainability methods, multimodal and longitudinal analysis, uncertainty estimation, privacy-preserving AI, and explainable foundation models for medical imaging.

Conclusion

This survey has consolidated the recent state of the art in deeplearningandExplainableArtificialIntelligenceasapplied to MRI-based brain tumour analysis. The reviewed evidence confirms that CNN-style architectures are very effective at extracting predictive features, but their black-box nature is the chief barrier to safe deployment in clinical practice. ArangeofXAItechniques—Grad-CAM,LRP,andSHAP beingthemostprominent—havebeenproposedtobridgethis gap.Theyprovidecomplementaryperspectives,fromregional saliency through pixel-level attribution to feature contribution analysis. Each, however, has its own granularity, computational overhead, and reliability trade-offs. Our principal observation is that no single explanation toolissufficientonitsown.Combiningseveralintoaunifiedframe-work yields multi-resolution rationales that are demonstrably more informative than any single output. This combination is especially valuable in healthcare, where transparency directly affects clinician trust and patient safety. The survey also flags several open challenges: there is still noconsensusonhowtoevaluateexplanationquality,several techniquesremaintoocostlyforreal-timeuse,andatranslation gappersistsbetweentheartefactsthatXAItoolsproduce and the language clinicians prefer. Tackling these issues — alongsidethefuturedirectionsoutlinedintheprevioussection is essential before such systems can move from research prototypes into routine radiology. In summary, unified XAI frameworks represent a practical step toward DL systems that are not only accurate but also interpretable,auditable,andclinicallyacceptable—threeprop-erties that, together, are required to make AI a trusted partner in medical decision-making. By coupling rigorous evaluation, lightweight implementations, multi-modal integration, human-centreddesign,robustnessanalysis,andsoundgovernance,the community can move steadily from promising prototypes to dependable clinical assistants.

References

[1] M.Z.KhalikiandM.S.Bas¸arslan,“Braintumordetectionfromimagesandcomparisonwithtransferlearningmethods,”ScientificReports,vol. 14, no. 1, 2024. [2] M. M. M. et al., “Enhancing brain tumor detection in MRI imagesthroughexplainableAIusingGrad-CAMwithResNet50,”BMCMedicalImaging, vol. 24, no. 1, 2024. [3] A. Al-Fakih et al., “FLAIR MRI sequence synthesis using squeezeattention generative model for reliable brain tumor segmentation,”Alexandria Engineering Journal, vol. 99, pp. 108–123, 2024. [4] P. Narayankar and V. P. Baligar, “Explainability of brain tumor classifi-cation based on region,” in Proc. Int. Conf. Emerging Technologies inComputer Science (ICETCS), 2024, pp. 1–6. [5] S.Ahmed,S.N.Nobel,andO.Ullah,“AneffectivedeepCNNmodel for multiclass brain tumor detection using MRI images andSHAP explainability,” in Proc. Int. Conf. Electrical, Computer andCommunication Engineering (ECCE), 2023, pp. 1–6. [6] T.HussainandH.Shouno,“Explainabledeeplearningapproachformulti-classbrainMRItumorclassificationandlocalizationusingGrad-CAM,”Information, vol. 14, no. 12, 2023. [7] F. S?efc?´?k and W. Benes?ova´, “Improving a neural network model byexplanation-guidedtrainingforgliomaclassificationbasedonMRIdata,”Int. J. Information Technology, vol. 15, no. 5, pp. 2593–2601, 2023. [8] M. I. Mahmud, M. Mamun, and A. Abdelgawad, “A deep analysis ofbrain tumor detection from MR images using deep learning networks,”Algorithms, vol. 16, no. 4, 2023. [9] E. M. Senan et al., “Early diagnosis of brain tumour MRI images usinghybrid techniques between deep and machine learning,” Computationaland Mathematical Methods in Medicine, 2022. [10] R. Najjar, “Redefining radiology: A review of artificial intelligenceintegration in medical imaging,” Diagnostics, vol. 13, no. 17, 2023.

Copyright

Copyright © 2026 Kamalakshi Naganna, Ankith V Hullamani, Aaryan Kumar, Tejaswini K, Yashwanth E S. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.