SIM-FISH: Diffusion-based simulation for fluorescence image in situ hybridization

Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH) is a powerful imaging technology for spatial transcriptomics, enabling the detection of numerous RNA species in individual cells and their spatial localization within the sampled region. In MERFISH, each RNA molecule is encoded by a unique binary barcode composed of multiple bits, where each bit corresponds to a single round of fluorescence imaging capturing the presence or absence of fluorescence signals for all RNA molecules at that position. By combining these sequential imaging rounds and decoding the bit patterns according to a predefined codebook, MERFISH achieves precise RNA identification and localization. However, MERFISH experiments are costly, time-consuming, and require live tissue samples maintained under tightly controlled culture conditions. These practical limitations make large-scale experimentation difficult. However, cost-effective computational approaches for spatial transcriptomics remain insufficiently explored. As a result, simulating MERFISH images becomes crucial for benchmarking new computational methods and exploring experimental designs without the overhead of physical imaging. In this study, we present a Stable Diffusion model enhanced with ControlNet (SD-CN) for synthesizing fluorescence bit-images that preserve RNA spatial information. Our methodology introduces three key innovations: 1. Transfer learning from pre-trained models to overcome data scarcity in fluorescence imaging. 2. Channel-specific Low-Rank Adaptation (LoRA) for fine-tuning the model to differentiate fluorescence patterns for each bit accurately. 3. ControlNet integration for enforcing spatial constraints using a spatial map with cell boundaries and cell types, ensuring biologically realistic spatial arrangements. We evaluated three variants of the Stable Diffusion model: (i) the vanilla Stable Diffusion model, (ii) a LoRA-adapted Stable Diffusion model without additional guidance, and (iii) the proposed SD-CN framework, which integrates ControlNet-based conditioning. Our results demonstrate that the SD-CN framework outperforms the other variants on key metrics, including texture similarity, color consistency, and spatial location similarity. This study explored SD-CN as a computational method to simulate MERFISH images. The model facilitates the development an validation of computational methods for spatial transcriptomics.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Wu, Ruochen

Contributors dc:contributor

• Shomorony, Ilan

• Stable Diffusion
• Controlnet
• Merfish
• Fluorescence

Statement dc:rights

• Copyright 2025 Ruochen Wu

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129625

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129625