Super-resolution chemical imaging via a diffusion-based deep generative model

Project Summary/Abstract The overarching objective of this project is to develop an advanced computational super-resolution model to enhance the throughput and resolution of chemical imaging. Chemical imaging, which includes advanced techniques such as stimulated Raman scattering (SRS) microscopy, holds immense potential in intraoperative

cancer detection. This is due to its ability to generate intrinsic molecular contrasts without tissue processing or labeling, which can improve the accuracy and speed of intraoperative diagnosis. Such improvement is crucial to better patient outcomes by providing near real-time feedback to the surgeon and reducing the risk of leftover

cancerous tissues. However, existing chemical imaging techniques grapple with an inherent limitation – the tradeoff between spatial resolution and imaging field of view, resulting in low imaging throughput and prolonged imaging durations for larger tissue samples. For other applications involving live cell imaging, the limited

resolution of SRS (> 300nm) hinders visualization of subcellular organelles and fine structures. Computational super-resolution, bolstered by advancements in artificial intelligence, can address these challenges by transforming low-resolution images into high-resolution versions. This has been achieved with the

Convolutional Neural Network and the Generative Adversarial Network. However, super-resolution chemical imaging is scarce. There are no datasets available for super-resolution training. It is also unclear whether existing super-resolution microscopy techniques could work for chemical images due to vast differences in imaging

contrasts. Here we propose to develop a new super-resolution technique ChemDiffuse that is based on the diffusion-based deep generative network. Diffusion-based models are widely used in popular image-generation tools such as Midjourney and DALL-E. While superior in stability and image quality to CNN and GAN models,

they need extensive training data. Leveraging on our recent progress in image augmentation and a new diffusion model for 2D and 3D data, we will develop the ChemDiffuse model to significantly improve SRS imaging throughput and resolution. We aim to test the application of the ChemDiffuse super-resolution model in two

different areas: 1. Fast gigapixel SRS imaging of tissue at submicron resolution for pathology application. We will use mouse brain tissue as our test system to train the ChemDiffuse model to enable fast 3D SRS imaging at 10 million pixels/sec, a 40-fold improvement in lateral dimension, and another 10-fold improvement in axial

dimensional. Such improvement is crucial for intraoperative stimulated Raman histology of large tissues. 2. Label-free SRS imaging of live cell organelles at 150 nm resolution. We will train the super-resolution model to enable 2-fold resolution enhancement with regular SRS imaging, an improvement that will allow unprecedented

label-free tracking of multiple organelles and single cell analysis for a wide range of drug discovery applications. Though our focus is on SRS imaging, this methodology can benefit various other chemical imaging techniques, including Raman microscopy, IR imaging, transient absorption microscopy, and photothermal microscopy, etc.