Medical Image Generation using Denoising Diffusion Probabilistic Model

Denoising Diffusion Probabilistic Models (DDPMs) have emerged as a new state of the art for image generation. Unlike previous approaches, like GANs and VAEs (which sometimes have difficulty during training or suffer from issues such as mode collapse), DDPMs follow a more stable approach that gradually diminishes noise in an image over many steps. While this method is less efficient, it tends to yield cleaner and more realistic results. In this work, we propose a class-conditional DDPM for synthesizing skin lesion images with a curated Skin Diseases Dataset. The image data consists of three lesion categories: benign keratosis-like lesions (BKL), melanocytic nevi (NV) and vascular lesions (VASC). Our model leverages a U-Net to predict the noise added by time-point in diffusion and takes class input to enforce that the generated images belong to the desired lesion category. Training is performed with the usual DDPM objective, based on mean squared error between predicted noise and actual noise under a fixed noise-averaging schedule. The results demonstrate that our model can effectively synthesized images for all the three classes, and simulated appearance-alike visual characteristics fit to the evolutionary habit and manifestation of various types of lesions. To investigate how useful these synthetic images are for various data generation beyond s generation, we also analyze their effect on a downstream classification task. We form several variants of real images and DDPM-generated ones and train another classifier to investigate whether this type of augmentation can enhance diagnosis. In this work, we show both the generative capabilities of our model as well as the benefits of introducing synthetic data to medical image classification workflows.