Rule-Based Synthesis of Microscopy Images by Diffusion Refinement.

Deep learning methods in medical imaging often suffer from the limited availability of high-quality annotated data, especially for rare conditions. This data scarcity is largely due to the need for domain expertise and the time-consuming process of data collection and annotation. Recent advances in generative neural networks offer a promising solution by producing realistic synthetic images that can supplement or partially replace scarce real data. In this work, we propose a framework for synthesizing realistic microscopy images together with their corresponding structural annotations. The proposed method combines a procedural generator that encodes expert-defined diagnostic rules with a diffusion-based refinement that enhances visual realism while preserving the prescribed structure. We further introduce a multi-stage diffusion-based refinement process that utilizes a segmentation mask to guide refinement and ensure a predefined structure. We demonstrate the ability of the proposed framework on the use case of synthesizing microscopic images of motile cilia cross-sections, which are important for the diagnosis of Primary Ciliary Dyskinesia (PCD). Our results show that data created by the proposed approach can serve as both a complement to and a substitute for real training data in a downstream segmentation task.