NeuroAttnFuseNet dual branch attention fusion for brain tumor MRI classification

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Brain tumor MRI classification remains challenging because tumor appearance varies across cases, class boundaries can be visually subtle, and public 2D MRI datasets may contain redundancy or source-specific biases that inflate reported performance. This study proposes NeuroAttnFuseNet, a dual-path attention-guided fusion network for four-class MRI image classification into glioma, meningioma, pituitary, and no tumor. The model combines a Swin-T branch for hierarchical local morphological representation with a frozen DINOv2 ViT-B/14 branch that provides a 768-dimensional global descriptor from the [CLS] token. The Swin and DINOv2 features are concatenated and passed through a channel-attention gate and fusion MLP before final softmax classification, allowing local and global cues to be reweighted before prediction. Experiments used a merged public dataset from Figshare, SARTAJ, and Br35H containing 7023 images. Evaluation followed an image-level leakage-mitigated protocol with duplicate and near-duplicate screening, training-only augmentation, and a fixed hold-out test set of 1311 images. On this hold-out set, NeuroAttnFuseNet achieved accuracy 0.9924, macro precision 0.9922, macro recall 0.9917, macro F1 0.9919, macro AUC 0.9991, and macro average precision 0.9969, with 1301 correct predictions and 10 errors. Repeated five-fold evaluation with three random seeds gave 99.07 ± 0.10% accuracy, outperforming ConvNeXt, CoAtNet, Swin Transformer, MobileNetV2, InceptionV3, and ResNet50 under the same protocol. Ablation analysis showed that the channel-attention fusion improved five-fold accuracy by 0.26 percentage points over concatenation with MLP fusion. These findings indicate high image-level performance on the merged dataset, but they should not be interpreted as evidence of patient-level or external clinical generalization because patient identifiers and independent multi-center validation were unavailable.