Multimodal Medical Image Generation and Classification via Generative Adversarial Networks
Y Zhang · 2025
Medical image analysis is essential for accurate disease diagnosis, with multimodal imaging offering complementary information to enhance diagnostic precision. However, current GAN architectures exhibit limitations in feature fusion efficiency, synthesized image quality, and classification accuracy. In this study, a GAN-based multimodal medical image generation and classification model was constructed. First, features are extracted using multi-scale receptive fields, key channels are enhanced by combining the channel attention mechanism, and the cross-enhanced fusion module realizes the fusion of MRI and PET features. In addition, Generator 3D is responsible for generating realistic medical images and classifying diseases through the classification header. Discriminator 3D is used to discriminate the realism of the images. Meanwhile, a comprehensive loss function is designed to optimize the generator's performance by combining adversarial loss, correlation loss, mean loss, and classification loss. The experimental results show that the proposed method can generate high-quality medical images and perform well on the classification task. It effectively improves the performance of medical image analysis and provides strong support for clinical diagnosis.