The insufficiency and inadequate diversity of medical datasets have been significant problems in computer-aided diagnosis (CAD) systems and medical imaging research. Consequently, medical image synthesis has emerged as one of the most dynamic machine learning research areas presently. We present a novel approach for fast and precise semantic image synthesis using diffusion models. In particular, our method generates high-quality Lung CT images with precise lung nodule areas by utilizing the semantic labels and key features of the nodules in LIDC-IDRI dataset.
We use a diffusion model to learn the conditional distribution of the image pixels given the semantic label map, which enables us to generate realistic and diverse images by sampling from the learned distribution. To further improve the quality of the synthesis results, we evaluate several key pathology features of the lung nodules in LIDC-IDRI dataset and use classifier guidance with these key features to optimize the generated nodule areas.
We evaluate our method with various kinds of synthetic image quality metrics and show that it outperforms state-of-the-art methods in terms of visual quality, diversity, and fidelity to the input semantic labels. Our method has the potential to enable new applications in medical imaging, such as medical data augmentation, anomaly in-painting, and diagnosis training.