Synergistic transfer learning and adversarial networks for breast cancer diagnosis: benign vs. invasive classification

Abstract

Current breast cancer diagnosis methods often face limitations such as high cost, time consumption, and inter-observer variability. To address these challenges, this research proposes a novel deep learning framework that leverages generative adversarial networks (GANs) for data augmentation and transfer learning to enhance breast cancer classification using convolutional neural networks (CNNs). The framework uses a two-stage augmentation approach. First, a conditional Wasserstein GAN (cWGAN) generates synthetic breast cancer images based on clinical data, enhancing training stability and enabling targeted feature incorporation. Second, traditional augmentation techniques (e.g., rotation, flipping, cropping) are applied to both original and synthetic images. A multi-scale transfer learning technique is also employed, integrating three pre-trained CNNs (DenseNet-201, NasNetMobile, ResNet-101) with a multi-scale feature enrichment scheme, allowing the model to capture features at various scales. The framework was evaluated on the BreakHis dataset, achieving an accuracy of 99.2% for binary classification and 98.5% for multi-class classification, significantly outperforming existing methods. This framework offers a more efficient, cost-effective, and accurate approach for breast cancer diagnosis. Future work will focus on generalizing the framework to clinical datasets and integrating it into diagnostic workflows.

Keywords: Accuracy; Breast cancer diagnosis; Deep learning; Generative adversarial networks (GANs); Transfer learning.

Fig. 1

Proposed framework for classifying microscopic breast images. The workflow includes preprocessing with CGAN-based data augmentation, multiscale feature extraction using DenseNet-201, ResNet-101, and NesNet Mobile, and classification into benign or malignant categories. The model achieves high accuracy (99% binary, 98% multiclass) with robust diagnostic performance, outperforming traditional biopsy methods.

Fig. 2

Histopathology images from the BreakHis dataset at magnifications 40×, 100×, 200×, and 400×.

Fig. 3

Representative histopathology images from the BreakHis and ICIAR datasets.

Fig. 4

Proposed framework for breast cancer detection using advanced multiscale feature processing and fusion. (a) The framework integrates data augmentation, multiscale feature extraction using DenseNet-201, ResNet-101, and NesNet Mobile, and fine-tuning for precise tumor classification. (b) The Multiscale Contextual Feature module leverages residual blocks to capture contextual information effectively. (c) The Multi-Layer Feature Fusion module concatenates low-, middle-, and high-level features into a feature pyramid, ensuring feature balancing and enhancing detection accuracy. (d) The Chain Parallel Pooling module employs parallel pooling blocks to maintain feature complexity and robustness while improving the framework’s efficiency. This AI-powered framework enhances tumor detection with improved sensitivity and robustness (Details in).

Algorithm 1

Automated Breast Cancer Detection and Classification with cWGAN Data Augmentation

Fig. 5

BreakHis Training: Accuracy (a) steadily climbs, reaching 0.976 while validation stabilizes. Loss curves (b) decrease continuously, confirming convergence and optimal performance.

Fig. 6

ICIAR Training Progress: Both accuracy (a) and loss (b) curves trend favorably, showcasing model convergence and strong performance.

Fig. 7

Proposed framework maintains robust classification across BreakHis magnifications (40×–400×): consistent diagonal dominance in confusion matrices (a–d) confirms strong generalization.

Fig. 8

Proposed framework outperforms across feature extractors and magnifications (40x-400x) in BreakHis, as shown by consistent high AUC values in ROC curves.

Fig. 9

Proposed Framework Performance on ICIAR (a,b): Confusion matrix and ROC curve demonstrate accurate classification and strong discriminative power.

Fig. 10

Samples after training (generated Benign images).