Evaluating conventional and deep learning segmentation for fast X-ray CT porosity measurements of polymer laser sintered AM parts

Laser sintering is evolving towards a genuine manufacturing technique for volume production and mass customized products. However, variability in part quality has to be reduced further to enable its use for high demanding and critical end-use applications. High volume and mass customized manufacturing impose the need for a fast and flexible measurement instrument, to automatically assess the overall part quality, which is currently not available for the AM industry. XCT has shown to be an effective tool to measure the part quality, but the large acquisition time still obstructs the use of XCT for in-line quality inspections of laser sintered parts. Altering the XCT settings to decrease the total acquisition time influences the SNR and CNR of the reconstruction, introduces artefacts and directly influences the segmentation quality and feature analyses. To minimize the influence of the deteriorated image quality, deep learning segmentation algorithms are evaluated and compared with conventional segmentation and denoising algorithms on low quality XCT scans with reduced acquisition times. The segmentation quality is quantitatively investigated with the Jaccard index, probability of detection, pore size distributions and porosity values and a qualitative comparison is provided. An improved segmentation for low-quality XCT scans is obtained by using deep learning segmentation algorithms while preserving a high generalization of the segmentation algorithm on low-quality XCT scans with a wide SNR and CNR range.

Automated summary

Laser sintering is evolving as a manufacturing technique for volume production and mass customized products. However, reducing variability in part quality is still necessary for high demanding applications. This study evaluates deep learning segmentation algorithms for low-quality XCT scans and compares them with conventional methods, showing improved segmentation while maintaining a high generalization capability.