Funded projects

Reconstruction of low-dose CTs using accelerated generative models based on diffusion

Fondef Idea I+D ID23I10053

Currently, computed tomography (CT) is one of the modalities of choice for multiple pathologies due to its excellent diagnostic performance and speed of acquisition. However, it is also the radiological modality responsible for the highest irradiation from diagnostic studies, reaching radiation levels equivalent to between 6 months and 5.1 years of exposure to environmental radiation. Although radiation dose reduction strategies are implemented in most of the latest generation equipment, it is not always possible to effectively reduce patient irradiation, as it implies a decrease in image quality, which has negative consequences for diagnosis. On the other hand, the use of deep learning (DL) has spread in recent years in radiology. There are commercial products that apply DL to image noise reduction with promising results, but they still lose clinical performance when faced with aggressive radiation reductions.

In this context, there is an opportunity to develop a solution that, using state-of-the-art generative DL, reconstructs images acquired with a radiation dose reduction greater than 50% without compromising diagnostic performance. Recent scientific work on diffusion models applied to low-dose CT reconstruction means we start from already formulated technological concepts, a TRL 2 maturity level. Our goal at the end of the project is to reach TRL 4: a technology validated in the laboratory.

Low-dose CT reconstruction with diffusion models

Results from this project include AST-n, a fast sampling approach for low-dose CT reconstruction with diffusion models (MIRASOL–MICCAI 2025), and a single-step flow-matching denoiser (MIRASOL–MICCAI 2026) — see publications.

← Funded projects