Enhanced Deep Learning Pipeline for Breast Cancer Classification using Progressive Growing Conditional GAN and Cross-Attention Ensemble Learning

Abstract

Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, with early detection being critical for improving patient outcomes. A major obstacle in developing robust automated diagnostic systems is the scarcity of well-annotated mammographic datasets, exacerbated by class imbalance and patient privacy constraints. This paper presents a novel GAN-Enhanced Deep Learning Pipeline (GEDLP) that integrates a Progressive Growing Conditional GAN (PG-cGAN) for high-fidelity synthetic mammogram generation with a dual-branch ensemble classifier for improved breast cancer detection and classification. The proposed PG-cGAN synthesizes realistic, pathology-specific mammogram images across four classes-Normal, Benign, Malignant (BIRADS III), and Malignant (BIRADS IV/V)-conditioned on lesion type, density grade, and patient age meta-data. The augmented dataset is fed into a dual-branch ResNet-50/EfficientNet-B4 ensemble with a cross-attention fusion module. Extensive experiments on the CBIS-DDSM, INbreast, and VinDr-Mammo benchmarks demonstrate that GEDLP achieves a classification accuracy of 97.6%, AUC-ROC of 0.987, sensitivity of 96.8%, and specificity of 98.1%, outperforming 12 state-of-the-art methods by margins of 3.2–9.7%. Fréchet Inception Distance (FID) of 8.74 and Inception Score (IS) of 6.21 confirm the superior realism of synthesized images. Ablation studies validate each component's contribution. The pipeline is publicly available as a modular open-source toolkit, enabling clinical researchers to augment limited datasets and deploy explainable AI-driven breast cancer screening systems.