Breast cancer is one of the major causes of deaths in women. In the meantime, proper and early diagnosis with the help of mammograms can greatly enhance the outcomes of treatment. Nevertheless, the volume of the available data and the inconsistency of lesions are significant obstacles to the development of reliable diagnostic models. Generative Adversarial Networks (GANs) can provide a solution to data augmentation but due to the static nature of their models, they are incapable of capturing diagnostically important features, like irregular mass margins or microcalcifications. The proposed research study proposes a new Diagnostic-Driven Topology-Adaptive GAN (DTA-GAN) framework that improves the performance by adapting the generator and discriminator structures in real-time during the training process based on the diagnosis. DTA-GAN is compared with two baseline models, four state-of-the-art GAN-based models, and five ablation-based DTA-GAN models in 13 metrics on three popular datasets: CBIS-DDSM, INbreast, and Mini-MIAS to perform an extensive assessment. DTA-GANs results remarkably exceed the benchmarks with an AUC of 0.90–0.92 (an increase of 10% over DCGANs 0.82), an FID of 8.40–8.45 (compared to StyleGAN2’s 7.89), and feature preservation LMS metrics of 0.87–0.89 and CDR 91.9–92.7% using both qualitative and quantitative assessment across the three datasets. DTA-GAN synthesizes mammograms with topology controller reward functions focusing on imaging and diagnostics to improve the classification accuracy of the following tasks, offering a breast cancer detection solution that can be scaled to be used widely. This is a breakthrough in medical imaging because it synthesizes data that meets the stringent diagnostic standards, making the systems that are used to make diagnoses more reliable and more generalizable.
