A review of synthetic and augmented training data for machine learning in ultrasonic non-destructive evaluation.

Ultrasonic Testing (UT) has seen increasing application of machine learning (ML) in recent years, promoting higher-level automation and decision-making in flaw detection and classification. Building a generalized training dataset to apply ML in non-destructive evaluation (NDE), and thus UT, is exceptionally difficult since data on pristine and representative flawed specimens are needed. Yet, in most UT test cases flawed specimen data is inherently rare making data coverage the leading problem when applying ML. Common data augmentation (DA) strategies offer limited solutions as they don't increase the dataset variance, which can lead to overfitting of the training data. The virtual defect method and the recent application of generative adversarial neural networks (GANs) in UT are sophisticated DA methods targeting to solve this problem. On the other hand, well-established research in modeling ultrasonic wave propagations allows for the generation of synthetic UT training data. In this context, we present a first thematic review to summarize the progress of the last decades on synthetic and augmented UT training data in NDE. Additionally, an overview of methods for synthetic UT data generation and augmentation is presented. Among numerical methods such as finite element, finite difference, and elastodynamic finite integration methods, semi-analytical methods such as general point source synthesis, superposition of Gaussian beams, and the pencil method as well as other UT modeling software are presented and discussed. Likewise, existing DA methods for one- and multidimensional UT data, feature space augmentation, and GANs for augmentation are presented and discussed. The paper closes with an in-detail discussion of the advantages and limitations of existing methods for both synthetic UT training data generation and DA of UT data to aid the decision-making of the reader for the application to specific test cases.

Comparison of the Classification Results Accuracy for CT Soft Tissue and Bone Reconstructions in Detecting the Porosity of a Spongy Tissue.

The aim of the study was to compare the accuracy of the classification pertaining to the results of two types of soft tissue and bone reconstructions of the spinal CT in detecting the porosity of L1 vertebral body spongy tissue. The dataset for each type of reconstruction (high-resolution bone reconstruction and soft tissue reconstruction) included 400 sponge tissue images from 50 healthy patients and 50 patients with osteoporosis. Texture feature descriptors were calculated based on the statistical analysis of the grey image histogram, autoregression model, and wavelet transform. The data dimensional reduction was applied by feature selection using nine methods representing various approaches (filter, wrapper, and embedded methods). Eleven methods were used to build the classifier models. In the learning process, hyperparametric optimization based on the grid search method was applied. On this basis, the most effective model and the optimal subset of features for each selection method used were determined. In the case of bone reconstruction images, four models achieved a maximum accuracy of 92%, one of which had the highest sensitivity of 95%, with a specificity of 89%. For soft tissue reconstruction images, five models achieved the highest testing accuracy of 95%, whereas the other quality indices (TPR and TNR) were also equal to 95%. The research showed that the images derived from soft tissue reconstruction allow for obtaining more accurate values of texture parameters, which increases the accuracy of the classification and offers better possibilities for diagnosing osteoporosis.

Fractal analysis of the computed tomography images of vertebrae on the thoraco-lumbar region in diagnosing osteoporotic bone damage.

Fractal analysis was used in the study to determine a set of feature descriptors which could be applied in the process of diagnosing bone damage caused by osteoporosis. The subject of the research involved the computed tomography images of vertebrae on the thoraco-lumbar region. The data set contained the images of healthy patients and patients diagnosed with osteoporosis. On the basis of fractal analysis and feature selection by linear stepwise regression, three descriptors were obtained. They were two fractal dimensions calculated with the variation method (transect - first differences and filter 1 estimators) and one fractal lacunarity calculated by means of the box counting method. The first two descriptors were obtained as a result of the analysis of grey images, and the third was the result of analysis of binary images. The effectiveness of the descriptors was verified using six popular supervised classification methods: linear and quadratic discriminant analysis, naive Bayes classifier, decision tree, K-nearest neighbours and random forests. The best results were obtained using the K-nearest neighbours classifier; they were as follows: overall classification accuracy - 81%, classification sensitivity - 78%, classification specificity - 90%, positive predictive value - 90%, and negative predictive value - 77%. The results of the research showed that fractal analysis can be a useful tool to extract feature vector of spinal computed tomography images in the diagnosis of osteoporotic bone defects.

Electrostatically Driven In-Plane Silicon Micropump for Modular Configuration.

In this paper, an in-plane reciprocating displacement micropump for liquids and gases which is actuated by a new class of electrostatic bending actuators is reported. The so-called "Nano Electrostatic Drive" is capable of deflecting beyond the electrode gap distance, enabling large generated forces and deflections. Depending on the requirements of the targeted system, the micropump can be modularly designed to meet the specified differential pressures and flow rates by a serial and parallel arrangement of equally working pumping base units. Two selected, medium specific micropump test structure devices for pumping air and isopropanol were designed and investigated. An analytical approach of the driving unit is presented and two-way Fluid-Structure Interaction (FSI) simulations of the micropump were carried out to determine the dynamic behavior. The simulation showed that the test structure device designed for air expected to overcome a total differential pressure of 130 kPa and deliver a flow rate of 0.11 sccm at a 265 Hz driving frequency. The isopropanol design is expected to generate 210 kPa and pump 0.01 sccm at 21 Hz. The device is monolithically fabricated by CMOS-compatible bulk micromachining processes under the use of standard materials only, such as crystalline silicon, silicon dioxide and alumina.

Logistic regression model for identification of right ventricular dysfunction in patients with acute pulmonary embolism by means of computed tomography.

PURPOSE: Diagnosis of right ventricular dysfunction in patients with acute pulmonary embolism (PE) is known to be associated with increased risk of mortality. The aim of the study was to calculate a logistic regression model for reliable identification of right ventricular dysfunction (RVD) in patients diagnosed with computed tomography pulmonary angiography. MATERIAL AND METHODS: Ninety-seven consecutive patients with acute pulmonary embolism were divided into groups with and without RVD basing upon echocardiographic measurement of pulmonary artery systolic pressure (PASP). PE severity was graded with the pulmonary obstruction score. CT measurements of heart chambers and mediastinal vessels were performed; position of interventricular septum and presence of contrast reflux into the inferior vena cava were also recorded. The logistic regression model was prepared by means of stepwise logistic regression. RESULTS: Among the used parameters, the final model consisted of pulmonary obstruction score, short axis diameter of right ventricle and diameter of inferior vena cava. The calculated model is characterized by 79% sensitivity and 81% specificity, and its performance was significantly better than single CT-based measurements. CONCLUSION: Logistic regression model identifies RVD significantly better, than single CT-based measurements.

High resolution ultrasonography for imaging metacarpophalangeal joints.

The purpose of the study was to examine the possibilities of high resolution US to visualize and diagnose the changes within the fissures of hand metacarpophalangeal joints and to assess the dimensions of metacarpophalangeal joints in healthy people and selected diseases. The study involved 44 individuals subjected to ultrasound examinations of the hands. The group included 35 healthy right-handed persons (22 women and 13 men), aged 21-53 (average - 33), 7 right-handed patients with rheumatoid arthritis (RA) (5 women and 2 men), aged 41-73 (average - 54) and 2 right-handed men treated for acromegaly, aged 43 and 58. The Logic 500 ultrasonograph (GE) was used. The images of metacarpophalangeal joints of both hands in the dorsal transverse and longitudinal projections were performed using the broad-band linear head with the frequency of 8.5-11 MHz. The analysis included 880 measurements of the metacarpophalangeal joints. The results were presented using the metacarpophalangeal joints of the 1st and 3rd fingers. The results in men and women, healthy and ill individuals were compared. In the authors' opinion high resolution US is a valuable method of imaging the anatomical structures of metacarpophalangeal joints, which may be used to evaluate morphological changes and to estimate the degree of joint destruction in some diseases.