Adaptive deep Q-GAN framework for enhanced breast cancer detection in medical imaging

Breast cancer continues to be one of the most perilous diseases, necessitating precise and prompt detection for effective intervention. Despite the potential of deep learning approaches in mammogram-based diagnosis, they encounter ongoing obstacles such as dataset imbalance, data scarcity, subpar synthetic augmentation, and duplicate feature inclusion, all of which diminish detection performance and generalization. This article presents an Adaptive Deep Q-GAN architecture that integrates Deep Q-learning into the training of Generative Adversarial Networks, thereby mitigating mode collapse, stabilizing learning, and generating diverse, high-quality synthetic tumor images. U-Net segmentation is utilized to delineate specific regions of interest from mammography images, which undergo preprocessing involving normalization, median filtering, and histogram equalization. The Cuckoo Optimization Algorithm is employed for feature selection, discarding irrelevant and duplicated characteristics to diminish computing complexity. A CNN-based classifier, trained on both actual and synthetic data, achieves enhanced accuracy in tumor classification. The experimental assessment on the CBIS-DDSM dataset indicates that the suggested method attains an accuracy of 99.24%, surpassing leading techniques by as much as 6.8%, while also achieving a 32% improvement in computing efficiency. The framework necessitates substantial training time due to the incorporation of reinforcement learning; however, this constraint can be alleviated by parallelization techniques. The findings demonstrate that the suggested method provides a reliable, generalizable, and therapeutically pertinent solution for automated breast cancer diagnosis in medical imaging.