Work

A hybrid scale attention network for uncertainty aware brain tumor MRI classification

Year: 2026

Type: article

Abstract: Accurate and reliable classification of brain tumors from magnetic resonance imaging (MRI) is critical for timely clinical decision-making. Existing deep learning methods suffer from three key limitations: overconfident predictions without principled uncertainty estimates, lack of external validation across acquisition protocols, and parameter-heavy architectures unsuitable for resource-constrained deployment. To address these gaps, we propose Hybrid Scale-Attention Network, a lightweight framework comprising an EfficientNet-B3 backbone, a depthwise-separable Adaptive Multi-Scale Module (AMSM), and an Evidential Deep Learning (EDL) classification head. AMSM fuses three parallel dilated convolutional branches at rates {1, 2, 4} and an identity branch via learned softmax gating, forming a convex combination of multi-scale representations at an approximately 6.6× lower parameter cost than equivalent standard convolutions. The EDL head enables single-pass evidential uncertainty estimation based on Dirichlet strength. Through systematic four-variant ablation, we find that stacking a Dual Attention Module after AMSM degrades cross-dataset generalization despite modest in-domain gains, motivating its exclusion. On BRISC 2025, the proposed model achieves 99.20% accuracy and 99.30% macro-F1 with only 10.97 M parameters. Development-stage external validation on the PMRAM-BD cohort yields 97.41% accuracy without fine-tuning, indicating strong cross-dataset performance. Misclassified samples carry higher mean evidential uncertainty than correctly classified ones, with a misclassification detection AUROC of 0.9461 at single-pass inference cost (16.8 ms), offering an efficient alternative to MC-Dropout. Grad-CAM and UA-GradCAM visualizations alongside MSAD provide radiologist-interpretable localization and per-class scale-branch analysis. These results demonstrate that the proposed model offers a favorable trade-off among accuracy, parameter efficiency, and principled uncertainty quantification, supporting potential retrospective clinical investigation.

Source: Discover Artificial Intelligence

DOI: 10.1007/s44163-026-02386-7