X-ray ptychography is a type of imaging that can achieve nanoscale resolution by measuring multiple diffraction patterns in an object over an extended region. The challenge, however, is in distinguishing unwanted areas that are costly to reconstruct but of little interest. A common approach to reduce this unwanted data is to first run a coarse scan on a large field of view to estimate the position of the desired object and then run a finer scan on a smaller field of view. Unfortunately, such an approach can still include a significant number of unwanted data points.

Various techniques have been developed to extract scattering direction and object transmission information from the diffraction patterns. But these techniques, too, have limitations. Most significantly, although they can provide useful feedback about the object’s properties, the final reconstruction still requires human experts to identify the important data points.

To address this problem, the Argonne team turned to machine learning – specifically, unsupervised learning, which can identify diffraction patterns from unlabeled data.

“Our aim was to achieve automatic data selection without human intervention,” said Zichao (Wendy) Di, a computational mathematician with a joint appointment in Argonne’s Mathematics and Computer Science division and the X-ray Science division.

In their approach, the researchers used information about both scattering direction and absorption. Absorption contrast is effective primarily in identifying regions with high optical density, whereas the scattering orientation method is most effective for detecting scattered light changes in regions with feature changes. ​“By combining these in a multimodal approach, we are not constrained by the otherwise incomplete object information from the separate techniques,” Di said.

Using this combination, the Argonne team developed a physics-informed unsupervised learning algorithm to identify diffraction patterns at scan positions within the region of interest. Only the ​“important” diffraction patterns were used in the reconstruction of the image, thus saving computational resources.

But how good was the reconstruction quality when using only this selected data? To answer this question, the researchers compared the full dataset (with all the diffraction patterns) and various reconstructions with downselected diffraction patterns. The results showed that their approach was effective.

Another concern was how much time was saved. Again, the approach proved beneficial, requiring only about 20% of the time needed for the full reconstruction.

The research team also noted that the new approach provides another benefit as well. Features smaller than the beam size can be identified through preprocessing of the diffraction patterns by segmenting them; such features typically are hard to detect without advanced imaging techniques (see Fig. 1).

For the full article about this work, see Lin, D., Jiang, Y., Deng, J., and Di, Zichao Wendy. Unsupervised classification for region of interest in X-ray ptychography. Scientific Reports 13, 19747 (2023).
https://​www​.ncbi​.nlm​.nih​.gov/​p​m​c​/​a​r​t​i​c​l​e​s​/​P​M​C​1​0​6​4​3​5​5​3​/​p​d​f​/​4​1​5​9​8​_​2​0​2​3​_​A​r​t​i​c​l​e​_​4​5​3​3​6.pdf