Making Soft Materials Visible Using A Deep Learning Without Needing Perfect Data

What is it about?

This study introduces a deep learning method called Edge View Enhanced Phase Retrieval (EVEPR) to improve how we visualize soft and low-density materials like hydrogels using X-ray micro-CT scans. Traditional X-ray CT techniques struggle to clearly show soft materials because they don’t absorb X-rays well. A special type of imaging, called phase contrast imaging, helps by capturing how X-rays bend as they pass through materials. However, current methods either enhance edges but are too noisy, or smooth images too much and lose important details. To solve this, we combined the strengths of two image types: one that shows clear edges and another that gives better overall contrast. We trained a computer model (called a convolutional neural network) to learn from both types and generate improved images. The resulting EVEPR images had higher image quality and made it easier to separate and analyze materials like hydrogel scaffolds, both in lab settings and in animal tissue samples. This method allowed us to build a reliable database for further automated analysis and may help researchers studying soft tissues or materials in medicine, biology, and pharmaceuticals.

Featured Image

Photo by Pawel Czerwinski on Unsplash

Photo by Pawel Czerwinski on Unsplash

Why is it important?

To begin, EVEPR solves a critical problem in imaging soft or low-density materials. Conventional X-ray imaging techniques struggle to visualize soft tissues, hydrogels, or biomaterials because they don’t absorb X-rays strongly. This limits our ability to analyze these materials accurately, especially for applications like tissue engineering, nerve repair, and drug delivery. EVEPR provides a way to make these invisible structures clearly visible. EVEPR also improves both image quality by measuring the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) while also providing consistent grey values. This is crucial for quantitative analysis and segmentation, enabling researchers to measure, model, or track materials accurately, without the inconsistencies or artifacts found in older methods. Perhaps the most important is that EVEPR belongs to a category of deep learning called , "self-supervised" deep learning which does not need an ideal ground truth. Most AI-based methods require "clean" and "high-quality" reference images, which are not always available. EVEPR works without a priori reference data by training on two types of imperfect images that contain complementary information. This makes the technique more versatile and scalable for real-world imaging tasks.

Perspectives