Scan Quality Estimation for Industrial Computed Tomography Using Convolutional Neural Networks

Artefacts in industrial Computed Tomography (CT) restrict the image quality of a CT scan and limit evaluations such as inspections for material defects or dimensional measurements. Due to a large variety of scanning objects made of different materials and of various part sizes, artefacts appear in various ways in the reconstructed image. Existing analytical approaches allow quantifying the CT scan quality, but still a lack of generalizability exists. In our work, a scan quality estimation based on a Convolutional Neural Network (CNN) is proposed. The network is trained with projection images of various objects. The images are labeled by pair-wise comparison by an experienced user with a scalar quality value. The training is performed on 94 scans covering a broad range of objects. Our work showed a test accuracy of the final trained network of up to 95 %. In order to comprehend the "black box" CNN, visualizations of the network's feature maps are analyzed.