A hybrid ConvNeXt-ViT framework with differential evolution optimization for breast cancer classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Automated breast cancer detection via mammography is crucial for early diagnosis but faces challenges like class imbalance and the need for both local and global feature extraction.
• Convolutional Neural Networks (CNNs) excel at local features but miss long-range dependencies, while Vision Transformers (ViTs) capture global context but require extensive data and computation.
• Existing methods struggle to balance efficiency and effectiveness in mammogram classification, hindering widespread clinical adoption.

Purpose of the Study:

• To introduce DEViTNeXt, a novel hybrid framework integrating ConvNeXt and Vision Transformer (ViT) with Differential Evolution (DE) optimization for improved mammogram classification.
• To address limitations of existing models in extracting both local and global features while managing dataset constraints and computational demands.
• To enhance the accuracy and reliability of automated breast cancer detection systems.

Main Methods:

• DEViTNeXt employs a dual-branch architecture: ConvNeXt for hierarchical local features and ViT for global context, refined by Multi-Head Attention (MHA).
• Preprocessing includes Gaussian filtering and CLAHE enhancement, coupled with a hybrid augmentation strategy using Generative Adversarial Networks (GANs) and geometric transformations.
• A DE-optimized MHA fusion layer adaptively integrates features, and a composite loss function (Weighted Cross-Entropy + Focal Loss) tackles class imbalance and difficult cases.

Main Results:

• DEViTNeXt achieved superior performance on the CBIS-DDSM dataset, reaching 99.63% accuracy, 99.45% sensitivity, and 99.55% specificity in binary classification.
• On the MIAS dataset, the model attained 98.50% accuracy for 3-class classification (Benign, Malignant, Normal).
• The proposed framework outperformed existing state-of-the-art methods in mammogram classification tasks.

Conclusions:

• DEViTNeXt offers a powerful and efficient hybrid approach for automated breast cancer detection, effectively combining local and global feature extraction.
• The model's high accuracy and sensitivity demonstrate its potential for improving early breast cancer diagnosis and patient outcomes.
• DEViTNeXt represents a significant advancement in medical imaging AI, overcoming limitations of previous CNN and ViT-based methods.