C2D-TBBA-Net: Clinically-Conditioned Diffusion with Tumor-Boundary Aware Attention Network for Real-Time Multi-Class Brain Tumor Detection on Edge Devices

Abstract

Brain tumor classification from magnetic resonance imaging (MRI) remains challenging due to limited annotated datasets, subtle boundary characteristics between tumor types, and the need for efficient diagnostic tools. We propose C2D-TBBA-Net, a novel framework combining clinically-conditioned diffusion augmentation with a Tumor-Boundary-Aware Attention (TBAA) mechanism for efficient multi-class brain tumor detection. Our conditioned diffusion model generates synthetic MRI images controlled by five tumor-specific parameters: type, size, location, boundary characteristics, and intensity profile, enabling targeted generation of clinically-relevant cases. The TBAA module explicitly models irregular tumor margins through multi-channel depthwise edge detection with learned channel-wise importance weighting, boundary-weighted spatial attention, and edge-aware channel attention. Our ultra-lightweight architecture achieves 97.8±0.3% accuracy on the Figshare dataset (n=613) and 94.7±0.7% on external Kaggle validation (n=1,311) with only 2.4M parameters, demonstrating statistical significance over seven baseline methods (p<0.001). The model maintains inference times of 24ms on mobile GPU (Snapdragon 8 Gen 2), making it suitable for point-of-care deployment. Cross-dataset validation demonstrates superior generalization with only 3.1% accuracy drop compared to 5.2-6.3% for baseline methods. Clinically-relevant attention focus is confirmed by radiologist evaluation (κ=0.82, IoU=0.79). Although the outcomes in resource-constrained environments are encouraging, future clinical validation is needed before application in the field.

Keywords

Brain Tumor Classification, Conditioned Diffusion Models, Boundary-aware Attention, Lightweight Deep Learning, Edge Computing, Mobile Healthcare