Relational Explainability of Brain Tumor MRI: Linking Classification Attention to Tumor Masks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Attention-based explanations (e.g., Grad-CAM++ (Gradient-weighted Class Activation Mapping Plus)) are commonly used in medical imaging, but their anatomical validity is rarely evaluated quantitatively. In this work we propose a novel relational explainability framework which goes beyond the conventional qualitative attention visualization by quantitatively validating the spatial correspondence between the classifier attention and the segmented tumor regions. Brain tumor segmentation is performed using DeepLabV3+ with EfficientNet-B4 encoder and tumor classification is performed using EfficientNet-B3. The Grad-CAM++ is employed to produce class-discriminative attention maps and the Exact Top-20% attention region is derived for quantitative analysis. The framework is evaluated on brain MRI datasets of gliomas, meningiomas, pituitary tumors, and non-tumor cases. The proposed method achieved a Dice coefficient of 87.5% for tumor segmentation and classification accuracy of 99.3% on the BRISC 2025 dataset. The classification accuracy on external evaluation was 93.5%. The quantitative validation of the explainability was performed by comparing the Exact Top-20% Grad-CAM++ attention region with the segmented tumor region in terms of Relational IoU, Tumor Coverage, and CAM Precision. The proposed framework achieved an overall Relational IoU of 0.0712, Tumor Coverage of 77.09%, and CAM Precision of 7.16%. These findings demonstrate class-dependent spatial alignment between classification attention and delineated tumor regions, with substantial tumor coverage but markedly lower spatial precision, underscoring the complementary rather than identical roles of discriminative attention and anatomical segmentation. The framework also identified the complementary functions of classification attention and anatomical segmentation, offering a more comprehensive understanding of the model’s behaviour. The proposed framework is an architecture-independent validation strategy providing an objective and quantitative approach to evaluate attention-based explanations, contributing to improving the transparency, interpretability and trustworthiness of artificial intelligence systems for medical image analysis.