BreasTransNeXt: An Enhanced Multi-Module Vision Transformer For Early Breast Cancer Diagnosis

Breast cancer (BC) represents the most common and one of the most lethal malignancies among women globally, ranking second only to lung cancer in terms of mortality. Early diagnosis is critically important, as treatment outcomes are significantly improved when BC is identified at an initial stage. Although considerable research has been conducted on BC classification utilizing deep learning techniques, there remains a deficiency of comprehensive models that effectively integrate robust feature extraction with balanced learning methodologies. To address this limitation, we introduce BreasTransNeXt, a new hybrid deep learning framework that synergizes the convolutional inductive bias inherent in ConvFormer with the efficient global attention mechanisms of a hybrid transformer architecture, thereby facilitating robust multi-scale feature representation. Furthermore, to mitigate class imbalance and enhance model generalization, the BC dataset is augmented through the application of a Wasserstein Generative Adversarial Network with gradient penalty, which generates high-fidelity synthetic images for both malignant and healthy classes. Additionally, BreasTransNeXt incorporates a hybrid attention mechanism comprising of local, channel, and global branches, which are integrated via adaptive gating; this design enhances feature representation by capturing fine-grained lesion characteristics, inter-feature correlations, and long-range contextual dependencies. The proposed model is rigorously evaluated against traditional convolutional neural network approaches and state-of-the-art vision transformer architectures. Experimental findings demonstrate that BreasTransNeXt consistently outperforms these methods, achieving an accuracy of 0.9710, recall of 0.9808, precision of 0.9808, and F1-score of 0.9808, thereby underscoring its potential as a reliable tool for early BC detection.