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

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Breast cancer (BC) is a leading cause of cancer mortality in women worldwide, making early diagnosis crucial for improved treatment outcomes.
• Existing deep learning models for BC classification often lack comprehensive feature extraction and balanced learning, hindering diagnostic accuracy.
• There is a need for advanced AI frameworks that can effectively integrate multi-scale feature representation and address class imbalance in BC datasets.

Purpose of the Study:

• To introduce BreasTransNeXt, a novel hybrid deep learning framework designed for robust and accurate breast cancer classification.
• To enhance feature representation by synergizing convolutional and transformer architectures for multi-scale analysis of BC images.
• To address the challenge of class imbalance in BC datasets using advanced data augmentation techniques.

Main Methods:

• Developed BreasTransNeXt, a hybrid deep learning model combining ConvFormer's convolutional inductive bias with a transformer's global attention mechanisms.
• Employed Wasserstein Generative Adversarial Network with gradient penalty for data augmentation, generating high-fidelity synthetic images to balance malignant and healthy classes.
• Integrated a hybrid attention mechanism (local, channel, global branches with adaptive gating) to capture fine-grained lesion characteristics and long-range dependencies.

Main Results:

• BreasTransNeXt achieved superior performance compared to traditional Convolutional Neural Network (CNN) and Vision Transformer (ViT) models.
• The model demonstrated high accuracy (0.9710), recall (0.9808), precision (0.9808), and F1-score (0.9808) in breast cancer classification.
• Experimental results validate the effectiveness of the hybrid architecture and attention mechanisms in enhancing feature representation and model generalization.

Conclusions:

• BreasTransNeXt offers a significant advancement in deep learning for breast cancer detection, providing robust multi-scale feature representation and effective class imbalance mitigation.
• The proposed framework shows strong potential as a reliable tool for early and accurate breast cancer diagnosis, improving upon existing AI methodologies.
• Further research can explore the clinical integration of BreasTransNeXt to aid radiologists and improve patient outcomes in breast cancer screening.