AI-based Feature Detection in X-ray-CT Images Using Synthesized Data

Matthew Konnik, Bahar Ahmadi, Nicholas May, Joseph Favata, Zahra Shahbazi, Sina Shahbazmohamadi, Pouya Tavousi
2020 Microscopy and Microanalysis  
Nondestructive volumetric analysis of samples, enabled by X-ray computed tomography (CT), has attracted scientists and engineers from a wide spectrum of disciplines that are interested in identification and measurement of miniature internal features of their samples [1] . While obtaining X-ray CT images of arbitrary objects has become a straightforward procedure, which only requires adjustment of a few imaging parameters (e.g., energy, # of projections), the interpretation of the resulting 3D
more » ... ages is still a challenging task [2] . For proper interpretation of an X-ray CT image, one must be able to extract welldefined geometric features from the raw data, where the raw data is a gray-scale 3D array of voxels [3] . Conventionally this task is performed manually by the subject matter experts (SMEs). Th extensive time and effort as well as human error associated with manual processes call for automated methods that can extract features accurately and with a high-throughput. The most common approach for achieving this goal is use of computer-vison (CV) techniques, to segment the images into distinct partitions [6] [7] [8] [9] [10] [11] [12] [13] [14], which could hopefully be used for extracting meaningful geometric features. For example, in thresholding, a common CV technique, intensity values and a preset thresholding constant will be used to assign a label to each pixel (voxel) in the 2D (3D) image. Such label is shared among all the pixels (voxels) of the same partition and the result of the segmentation process is a 2D (3D) image that is partitioned into several groups of connected pixels (voxels). Although, the CV techniques may offer an automated process in the absence of image noise (i.e., features that are of no interest), their performance drops drastically in dealing with noise which is prevalent in any image obtained from an X-ray CT practice [15] [16] [17]. The produced noise can be mitigated, but not completely removed. Therefore, in practice, the CV methods are only used to assist the manual feature extraction process and cannot provide a fully automated feature extraction process. The success of machine learning (ML) algorithms in automating tasks that are not analytically well-defined, promises use of these methods for automated feature extraction, as a superior alternative to CV-based methods. The idea is to train a machine learning algorithm with sufficient ground truth data [18] [19] [20] and then use it for automated feature extraction. Here, the ground truth data are obtained from labeled X-ray CT images. Each data point consists of: (1) raw data in the form of a gray-scale 3D array of voxels and (2) the corresponding feature. The caveat is that the proper training of a machine learning algorithm demands huge amounts of labeled data. This is a multifold challenge. The necessity of labeling the raw data manually, makes this process extremely tedious, if not impractical. In addition, labeling process will be subjective, with different outcomes expected from different SMEs. Further, such manual process is subject to error. Double-checking and triple-checking practices to eliminate such error would add to
doi:10.1017/s1431927620020498 fatcat:kykpeyd3ijgknn55cm46av6xlm