Toward interpretable multiclass semantic intelligence: explainable brain tumor segmentation and automated grade stratification

Abstract

Abstract Purpose Accurate brain tumor grade classification from magnetic resonance imaging is essential for diagnosis, treatment planning, and prognosis assessment. Although deep learning models have demonstrated strong performance, many lack clinical validation, robustness across datasets, and interpretability for decision support. This study proposes an explainable hybrid deep learning framework that integrates tumor segmentation, feature optimization, grade classification, and model attribution for high-grade and low-grade gliomas, with evaluation across multiple MRI datasets under controlled experimental conditions. Methods The proposed framework, Towards Interpretable Multiclass Semantic Intelligence: Explainable Brain Tumor Segmentation and Automated Grade Stratifi cation, integrates brain tumor segmentation, optimized feature selection, and discriminative classifi cation to support automated tumor analysis and grade stratifi cation. Tumor regions are segmented using a Path Aggregation Network (PANet) integrated with Graph Convolutional Neural Networks (GCNNs) and ResNet-50 to preserve relevant structural characteristics. From the segmented tumor regions, shape, texture, and Histogram of Oriented Gradients (HOG) features are extracted to characterize morphological and visual patterns. A modified Reptile Search Algorithm (RSA) is then employed to select discriminative features, aiming to reduce feature redundancy and mitigate over fitting. The selected features are subsequently classified using DenseNet-201 within a Siamese architecture to enhance robustness to inter-patient variability. Model interpretability is incorporated through Gradient-weighted Class Activation Mapping (Grad-CAM), which provides visual explanations of the regions contributing to model predictions. The framework is evaluated using the BraTS 2019, BraTS 2020, and BraTS 2021 datasets, together with additional MRI datasets. Results The proposed model achieved classification accuracies of 99.12%, 99.23%, and 98.86% on the benchmark datasets. The average precision, sensitivity, and specificity exceeded 98% across datasets, demonstrating strong discrimination between tumor grades and consistent generalization across imaging protocols. Visualization analysis highlighted tumor regions contributing to classification decisions. Conclusions The proposed model achieved classification accuracies of 99.12%, 99.23%, and 98.86% on the benchmark datasets. The average precision, sensitivity, and specificity exceeded 98% across datasets, demonstrating strong discrimination between tumor grades and consistent generalization across imaging protocols. Visualization analysis highlighted tumor regions contributing to classification decisions. Conclusions: The proposed framework demonstrates promising performance and provides interpretable model-attribution information under the evaluated experimental conditions. It may serve as a potential computer-aided decision-support approach, but prospective multicenter validation, calibration, clinician-in-the-loop assessment, and workflow integration are required before routine clinical deployment.