An efficient method for performing segmentation analysis, a crucial step in materials science, has been created by researchers at RIKEN, Japan’s synchrotron radiation facility, SPring-8, and their partners.
To comprehend a material’s fine-scale composition, segmentation analysis is utilized. It designates discrete areas (also known as “segments”) with particular compositions, structural traits, or qualities. This makes a material’s suitability for a given purpose and any potential drawbacks easier to assess. It can also be applied to the quality control of material fabrication and to the analysis of failed materials to pinpoint weak points in the material.
For synchrotron radiation X-ray computed tomography (SR-CT), which is similar to traditional medical CT scanning but uses intense focused X-rays created by electrons moving in a storage ring at nearly the speed of light, segmentation analysis is crucial.
In order to visualize the three-dimensional structure of samples with slight density differences between regions of interest, like epoxy resins, the team has shown that machine learning is capable of performing the segmentation analysis for the refraction contrast CT.
According to first author Satoru Hamamoto, there hasn’t been a published general segmentation analysis technique for synchrotron radiation refraction contrast CT up to this point. Segmentation analysis has typically required trial and error by researchers, which has made it challenging for non-experts.
To precisely adapt to the segmentation analysis of SR-CTs, the team’s solution combined a transfer learning technique with machine learning techniques developed in the biomedical domains. Using the improved machine learning model, the quantity of training data required to achieve the desired outcomes was significantly decreased.
Satoru Hamamoto, the lead author, notes that no general segmentation analysis technique for synchrotron radiation refraction contrast CT has been published since. Experts have found segmentation analysis challenging because researchers have typically had to conduct it through trial and error.
The team’s solution was to fine-tune to the segmentation analysis of SR-CTs by combining a transfer learning technique with machine learning techniques proven in biomedical fields. Building the current machine learning model significantly decreased the quantity of training data required to achieve the desired outcomes.
Using an epoxy resin, the researchers were able to identify areas that had been formed by water through a proof-of-concept analysis. Their accomplishments imply that the method will be helpful for examining a variety of materials.
Through a proof-of-concept analysis, the researchers were able to identify areas in an epoxy resin that were actually formed by water. Their accomplishment indicates that a variety of materials can be analyzed with the technique.
The group intends to start offering segmentation analysis as a service to outside researchers through the recently operational SPring-8 data center in order to make this analysis method as widely and rapidly accessible as possible.