An Improved Pix2Pix GAN for Medical Image Generation

Generating missing modality medical images using deep learning is crucial for disease diagnosis, treatment planning, and medical education. With the rapid development of deep learning, Generative Adversarial Networks (GANs) have shown significant potential by producing realistic, high-quality images. However, GANs face challenges such as blurred organ edges, diffuse tissue boundaries, and failure to maintain the target modality’s style. This study improves the classic GANs model by introducing the Channel Attention Module (CBAM) into the generator’s Residual Block, enhancing feature extraction and encoding capabilities. We integrate the VGG16 network to enhance the depiction of image edge intricacies and introduce a histogram chi-square comparison loss to tackle the challenges associated with bone structure generation, stemming from data imbalance. To enhance image clarity, we include a gradient loss term in the loss function. Additionally, real CT images are input into the generator to help the model learn the target modality’s style features. Using Conditional GANs, mismatched data pairs are input into the discriminator with false labels to improve content matching sensitivity. Extensive experiments on public brain datasets demonstrate that our method significantly enhances GANs’s image generation capabilities, producing images that excel in both visual quality and quantitative metrics.