Generating Post-healing Images of Skin Diseases Based on an Adversarial Self-coding Generator

Goal: In recent years, the rapid development of computer vision and deep learning has driven the improvement of new technologies in medical imaging, while applications in various aspects have emerged. Regarding dermatology, which is an important branch of medical imaging technology, the techniques surrounding it are currently focused on lesion segmentation, identification, and classification of benign and malignant cells in pathological sections. we explore a post-healing image generation method based on dermatological diseases to fill the current situation where the field is relatively under-researched in terms of post-healing. Methods: In this work, the ISIC 2018 dataset is used as our object to be recovered, and the U-net neural network is used as the network model of the segmentation module to segment the healthy skin cell region and the focal cell region by setting the threshold region. The erosion module is used to obtain the area to be filled for this lesion, thus simulating the different stages of healing. Finally, Images are generated by a VAE/GAN-based adversarial autoencoder. Conclusion: Our proposed method can generate more realistic post-healing images of skin diseases compared with the VAE and GAN methods. Significance: This work explores the application of AI-aided diagnosis in the generation of post-healing images in medical imaging, which can help doctors and patients understand the healing process after surgery to a certain extent.