In this blog post, we showcase the usage of Dragonfly’s super-resolution capabilities. Super-resolution in X-ray CT involves taking low-resolution and high-resolution scans and creating a deep-learning model using this pair. The deep learning super-resolution model then takes an input of poor resolution scans to create superior image quality results – almost like the high-resolution images but without the need to scan at high resolution!

The application selected for this demo is the characterization of additively manufactured lattice structures. These complex-designed porous structures have many potential applications: For more about their applications, refer to the recent review paper [1].

Despite their many advantages, additive manufacturing of lattice structures has some inherent limits, and manufacturing defects can occur that require careful inspection and characterization. X-ray CT is well suited to this but is inherently limited to the object size, making it often necessary to accept a poor resolution. Imagine a large object with some small section of lattice material, for example. You can overcome this challenge with deep-learning super-resolution imaging. In the pictures below (fig. 1 +2), we show a high-resolution (4 μm voxel size) and low-resolution (16 μm voxel size) scan of the same lattice – at poor resolution, lots of surface detail is lost. The scans in this demo were made using a Comet Yxlon FF35 CT system. The sample itself is 6 mm across and was manufactured by Laser Powder Bed Fusion (L-PBF) using an EOS M290 printing system with Ti6Al4V powder.

Fig. 1 + 2: Examples of high-resolution and poor-resolution scans of a lattice structure (sample width is 6 mm for scale)

The lack of detail in the poor resolution scan is simply due to the larger voxel size, as is seen in the comparison of slice images in figures 3 + 4, again for the two resolutions. The images are intentionally shown with individual voxels rather than with interpolation to emphasize the difference in voxel size.

Fig. 3 + 4: Comparison of low and high resolution showing individual voxels/pixels.

The capability to train a neural network to take input and output as two images and create a network that transforms all similar poor-resolution images into high-resolution outputs is shown in the example below in figure 5, where the middle image is the super-resolution result. The signal-to-noise ratio (SNR) improvement is clearly visible in the image and quantified using Dragonfly to the following: the original poor resolution SNR is 1.19, the high-resolution scan is 1.48, and the super-resolution result is 1.47.

Fig. 5: Super-resolution model (middle) is possible by training using one high-resolution (right) and one poor-resolution dataset (left), which can get applied on any future scans of poor resolution.

While the detail visibility is still limited due to the original pixel size, the super-resolution result does allow better quantification of small features due to the image enhancement achieved in the process. Shown in the figures 6 to 8 are 3D images of porosity in one strut of the lattice, imaged and segmented using the poor-, super-, and high-resolution datasets. The segmentation here was performed with automatic Otsu thresholding. The super resolution is significantly better than the poor resolution, even though it is based on the same pixel size and data input, achieving almost the same detail as the high-resolution result.

Fig. 6 - 8: Individual lattice strut regions with porosity analyzed: top left the poor-resolution, top right the super-resolution, and bottom the high-resolution results.

After applying the super-resolution model to the whole lattice, the resulting porosity analysis comparison is shown below in the figures 9 to 11, in this case, further improved using a deep-learning segmentation model (please refer to our blog post from 5th April, 2023 ‘Comet’s AI solution to porosity measurement).

Figure 9 - 11: Full lattice porosity analysis applied to poor-resolution, super-resolution, and high-resolution datasets.

In conclusion, the best possible inspection and quantification of defects in additively manufactured lattice structures should get performed by using the highest possible acquisition resolution. But in case that is not possible, super-resolution performs significantly well and allows much better analyses than poor-resolution datasets. The benefits of the improved contrast using super-resolution deep learning models include the ability to scan larger objects but improve their image quality or to use two systems more efficiently (e.g., scan at high resolution once, then always use poor-resolution, lower-cost system for further scans). High-resolution scans often require a more careful setup and longer scan times, so there is also an efficiency advantage. Super resolution is only one of the many applications of AI to X-ray computed tomography and provides significant advantages!

[1] Du Plessis, A., Razavi, S.M.J., Benedetti, M., Murchio, S., Leary, M., Watson, M., Bhate, D. and Berto, F., 2022. Properties and applications of additively manufactured metallic cellular materials: A review. Progress in Materials Science, 125, p.100918.