Improving Breast Cancer Performance in CNN by Generating Synthetic Histopathological Images using GAN and Traditional Augmentation

Abstract

In the pursuit of more accurate cancer detection through breast cancer histopathology (BCH) images, Convolutional Neural Networks (CNNs) have emerged as promising tools. However, CNNs still face limitations, necessitating advancements in classification performance. This research addresses these challenges by harnessing the power of Generative Adversarial Networks (GANs) as data augmentation to optimize CNN models for BCH image classification. This paper addresses the proposed two-stage augmentation strategies based on GAN and the traditional method. The BreakHis dataset was employed to investigate the efficacy of GAN-based data augmentation. The research adopted a transfer learning approach, namely Inception-V3, and VGG16, and fine-tuned them with a single GAN and the two stages augmentation methods. The novel integration of GANs and traditional augmentation enhanced the training dataset, enabling the models to learn from a more diverse and extensive image distribution. Extensive trials demonstrated that the top-performing architecture, Inception-V3 + TradAug, attained a remarkable 97.12% accuracy with 0.1014 loss value, showcasing the effectiveness of the composition of GAN and traditional augmentation in optimizing BCH image classification. The two-stage integration of GANs, such as data augmentation and traditional augmentation, empowers CNN models to identify cancerous conditions accurately. This research signifies a significant step towards enhancing breast cancer classification through advanced AI-driven methodologies.