CASUS Institute Seminar, Nikita Deshpande, Student Assistant, Machine Learning for Infection and Disease, CASUS/HZDR

Abstract of the talk// Super-resolution microscopy (SRM) has emerged as a pivotal tool to transcend the limitations of traditional imaging, offering unprecedented resolution at the nanoscopic level. Traditional fluorescence microscopy faces limitations due to the Abbe diffraction limit, hindering the visualization of small structures at the molecular level and leading to an inverse problem. To overcome this constraint, conventional SRM techniques have been developed. Model-based methods, including deconvolution techniques and deep learning super-resolution (DLSR) networks, offer alternative approaches to enhance resolution but have limitations. This thesis investigates the utilization of denoising diffusion probabilistic models (DDPM), a category of deep generative models, in SRM by using the publically available microscopic low-resolution (LR) and high-resolution (HR) image-paired BioSR dataset. It also explores the impact of the noise schedule employed in the diffusion process and a way to predict it. DDPM is the focus point in this study, known for its remarkable generative abilities and stable training objectives. The research presents conditional DDPM (C-DDPM) and improved conditional DDPM (IDDPM) models as baselines, adapted from Google’s SR3 model and introduces a new network named schedule predictor (SP-DDPM). This method introduces a novel network for predicting the noise schedule. This predictive model is integrated and fine-tuned with the primary network, effectively eliminating any reliance on the noise schedule. The methods’ performance is evaluated based on generated super-resolution images using various metrics, including peak-signal-to-noise-ratio (PSNR), structural similarity index (SSIM), multiscale SSIM (MS-SSIM), normalized root mean squared error (NRMSE), and resolution. Results demonstrate the effectiveness of the proposed SP?DDPM approach, showing promising behaviour compared to other DLSR methods on the f-actin structure from the dataset. This work advances the SRM technique using DDPM, offering potential applications in diverse scientific fields.

Nikita Deshpande will be talking live in Görlitz. However, as the event is organized in a hybrid format that includes a videoconferencing tool by Zoom Inc., people interested in the topic have the chance to also join the talk remotely. Please ask for the login details via contact@casus.science.