From black box to glass box: Clinically grounded CNN interpretability in brain MRI via anatomically-aware layer-wise relevance propagation

Convolutional Neural Networks (CNNs) is helpful in detecting brain tumors on MRI with accuracy comparable to that of radiologists, but their black-box nature hinders in clinical adoption, as explainability is essential. Previous studies utilize generic saliency methods (e.g., Grad-CAM), yet often lack anatomical accuracy, quantitative clinical validation, or practical robustness. This study presents a different approach, a clinically based LRP framework that surpasses previous methodologies in three essential dimensions: anatomically conscious LRP decomposition. We suggest a changed version of the ε–αβ rule that adapts based on gradient variance for each layer. This improves localization fidelity in areas of heterogeneous tissue (for example, oedema vs. solid tumor). Clinician-in-the-loop validation: We measure the quality of the explanation using Pearson correlation (r = 0.87) and Intersection-over- Union (IoU = 0.79) against expert radiologist annotations, which is a level of clinical grounding that is not often seen in previous XAI studies. Integrated decision-support utility: In blinded reader studies (n = 5 radiologists ,LRP overlays enhanced diagnostic confidence by 32% (p < 0.001, Cohen’s d = 1.1) and decreased interpretation time by 23% (95% CI: [18%, 28%]). Our glass-box system usesa fine-tuned VGG-16 on 3,000 multi-sequence MRI scans (BraTS 2020 + clinical cohort) a keeps a high level of accuracy (94.2% test accuracy, 95% CI: [92.8%, 95.5%]). It also makes pixel-level heatmaps that line up with anatomical structures (for example, hippocampal infiltration and cortical disruption). Unlike bounding-box detectors (e.g., YOLOv7: 99.5% detection, but no pathology characterization), our method is important because it supports diagnostic reasoning instead of just localisation. This work directly addresses a significant deficiency: trustworthy explainability for clinicians. We end with a plan for how to integrate clinical data, which includes ways to comply with FDA SaMD rules and federated learning for validation across multiple institutions