In order to better understand failure and fracture of materials, observe microstructural evolution in real time and perform physics simulations on real structures, 3D imaging techniques must be used. ZEISS ZEN Intellesis is also capable of segmenting 3D data sets (in a compatible format, e.g. .txm, .tiff or .czi), such as those generated by X-ray microscopy.
Labelling can be applied to any or all slices through the 3D dataset. The model trains on all labelled slices and thus tine inputs can come from any labels on any slices. It should be noted that there are no feature vectors operating within 3D – the feature vectors are created on a slice-by-slice basis. However, this effect may be mitigated by analysing the data with different slice angles and taking averages.
Figure 6 ZEISS ZEN Intellesis segmentation of a foam glass. (A) Virtual 2D slide of X-ray micrograph data set. (B) Segmented microstructure showing pores in blue, glass walls in red. (C) 3D model of the foam glass, using segmentation results. (D) Image of a typical foam glass structure. Sample courtesy of Martin Bonderup Østergaard, Dr. Rasmus R. Petersen and Prof. Yuanzheng Yue from Aalborg University, and Dr. Jakob König from Jozef Stefan Institute
In example of 3D data set segmentation is shown in Figure 6. The specimen is a foam glass insulator used in the construction industry. Researchers are interested in determining the porosity and internal structure of these materials, improving the synthesis process of mixing glass powder plus foaming agents and simulating its thermal properties using physics simulations on real 3D structures. In order to extract this information, obtain homogeneous pore size distribution, minimize defects and increase insulation capability, image segmentation in 3D of raw data is needed. Then, ZEISS ZEN Intellesis was able to create a 3D representation of the sample. To do this, a model was trained to segment the internal structure such that both large pores and smaller pores present in the glass walls are identified to produce accurate results.
Comparison of classical image segmentation algorithms (global multi-Otsu thresholding or seeded watershed growing) with machine learning multivariant classification in ZEISS ZEN Intellesis has also been carried out on synthetic images produced from actual 3D data sets. All algorithms performed well under low noise levels but machine learning classification was much more noise tolerant than the other algorithms. Machine learning was also able to segment based on textural contrast, which the traditional algorithms were unable to do. The two traditional techniques achieved misclassification rates of above 50% in the textural contrast regions (at zero noise levels), whereas for machine learning this dropped to below 5% misclassification.