A contrastive learning framework with adaptive feature fusion for brain tumor classification

Accurate classification of brain tumors from MRI scans is critical for diagnosis and treatment planning, yet remains challenging due to high intra-class heterogeneity and subtle inter-class differences. While deep learning has shown promise in automating this task, existing models often lack the discriminative power to reliably capture fine-grained pathological features. To address these limitations, we propose a novel contrastive learning framework with adaptive feature fusion (AFF-CL). Our method introduces a dynamic label queue that stores historical labels, enabling the construction of multiple positive pairs from different images of the same class. This design injects explicit supervision into contrastive learning, where the multiple positive pairs provide comprehensive intra-class variation coverage, directly facilitating feature representations that minimize intra-class distance while maximizing inter-class separation. Furthermore, to capture multi-scale contextual information, we augment the model with a local input stream and design an adaptive feature fusion (AFF) module to intelligently integrate localized features with global representations. Extensive experiments on the public figshare dataset demonstrate that our framework achieves state-of-the-art performance, significantly outperforming existing methods. The proposed approach shows strong potential for enhancing the precision of computer-aided diagnosis in clinical practice.