Application of machine learning classifiers to X-ray diffraction imaging with medically relevant phantoms

PURPOSE

Recent studies have demonstrated the ability to rapidly produce large field of view X-ray diffraction (XRD) images, which provide rich new data relevant to the understanding and analysis of disease. However, work has only just begun on developing algorithms that maximize the performance toward decision-making and diagnostic tasks. In this study, we present the implementation of and comparison between rules-based and machine learning (ML) classifiers on XRD images of medically relevant phantoms to explore the potential for increased classification performance.

METHODS

Medically relevant phantoms were utilized to provide well-characterized ground-truths for comparing classifier performance. Water and polylactic acid (PLA) plastic were used as surrogates for cancerous and healthy tissue, respectively, and phantoms were created with varying levels of spatial complexity and biologically relevant features for quantitative testing of classifier performance. Our previously developed X-ray scanner was used to acquire co-registered X-ray transmission and diffraction images of the phantoms. For classification algorithms, we explored and compared two rules-based classifiers (cross-correlation, or matched-filter, and linear least-squares unmixing) and two ML classifiers (support vector machines and shallow neural networks). Reference XRD spectra (measured by a commercial diffractometer) were provided to the rules-based algorithms, while 60% of the measured XRD pixels were used for training of the ML algorithms. The area under the receiver operating characteristic curve (AUC) was used as a comparative metric between the classification algorithms, along with the accuracy performance at the midpoint threshold for each classifier.

RESULTS

The AUC values for material classification were 0.994 (cross-correlation [CC]), 0.994 (least-squares [LS]), 0.995 (support vector machine [SVM]), and 0.999 (shallow neural network [SNN]). Setting the classification threshold to the midpoint for each classifier resulted in accuracy values of CC = 96.48%, LS = 96.48%, SVM = 97.36%, and SNN = 98.94%. If only considering pixels ±3 mm from water-PLA boundaries (where partial volume effects could occur due to imaging resolution limits), the classification accuracies were CC = 89.32%, LS = 89.32%, SVM = 92.03%, and SNN = 96.79%, demonstrating an even larger improvement produced by the machine-learned algorithms in spatial regions critical for imaging tasks. Classification by transmission data alone produced an AUC of 0.773 and accuracy of 85.45%, well below the performance levels of any of the classifiers applied to XRD image data.

CONCLUSIONS

We demonstrated that ML-based classifiers outperformed rules-based approaches in terms of overall classification accuracy and improved the spatially resolved classification performance on XRD images of medical phantoms. In particular, the ML algorithms demonstrated considerably improved performance whenever multiple materials existed in a single voxel. The quantitative performance gains demonstrate an avenue to extract and harness XRD imaging data to improve material analysis for research, industrial, and clinical applications.

Development and assessment of a multi-beam continuous-phantom-motion x-ray scatter projection imaging system

X-ray image formation using scattered radiation can yield a superior contrast-to-noise ratio compared to conventional transmission x-ray imaging. A barrier to practical implementation of scatter imaging systems has been slow image acquisition. We have developed a projection imaging system which uses five monoenergetic pencil beams in combination with continuous phantom motion to achieve acquisition times that are practical for medical and security applications. The system was configured at the Canadian Light Source synchrotron and consists of a primary collimator, motorized stages for phantom translation, a flat-panel x-ray detector for measuring scattered x rays, and photodiodes for simultaneously measuring transmitted x rays. Image generation requires several corrections to raw data artifacts arising from the nature of the detector, x-ray source, and acquisition procedure. We developed a novel correction for pixel location inaccuracy arising from continuous phantom motion. A five-beam system had nearly five times faster acquisition than a single-beam system. Continuous motion acquisition was approximately 30 times faster than step-and-shoot acquisition. The total acquisition time for a 9 cm × 5 cm phantom with 8425 pixels was just over 2 min. Image quality was also assessed, in part to determine its relation to acquisition speed. The width of sharp material boundaries was found to be at a minimum equal to the pencil beam width (1.75 mm) and to have an additional width equal to the product of the phantom translation speed and the acquisition time per pixel (up to 1.0 mm in our experiments). Contrast-detail performance was independent of acquisition speed, depending only on phantom entrance x-ray fluence. Pixel signal-to-noise ratio measurements indicate that detector readout noise is important for the scatter data, even for phantom air kerma as high as 30 mGy. Images could be improved with a detector having lower readout noise and higher sensitivity. Its spatial resolution could be moderate. We confirmed that for the same range of λ^-1 sin(θ/2), where λ is the x-ray wavelength and θ is the scattering angle, scatter images acquired using different beam energies (33-70 keV) had nearly identical contrast.

X-ray diffraction from bone employing annular and semi-annular beams

There is a compelling need for accurate, low cost diagnostics to identify osteo-tissues that are associated with a high risk of fracture within an individual. To satisfy this requirement the quantification of bone characteristics such as 'bone quality' need to exceed that provided currently by densitometry. Bone mineral chemistry and microstructure can be determined from coherent x-ray scatter signatures of bone specimens. Therefore, if these signatures can be measured, in vivo, to an appropriate accuracy it should be possible by extending terms within a fracture risk model to improve fracture risk prediction.In this preliminary study we present an examination of a new x-ray diffraction technique that employs hollow annular and semi-annular beams to measure aspects of 'bone quality'. We present diffractograms obtained with our approach from ex vivo bone specimens at Mo Kα and W Kα energies. Primary data is parameterized to provide estimates of bone characteristics and to indicate the precision with which these can be determined.

High energy transmission annular beam X-ray diffraction

We demonstrate material phase retrieval by linearly translating extended polycrystalline samples along the symmetry axis of an annular beam of high-energy X-rays. A series of pseudo-monochromatic diffraction images are recorded from the dark region encompassed by the beam. We measure Bragg maxima from different annular gauge volumes in the form of bright spots in the X-ray diffraction intensity. We present the experiment data from three materials with different crystallographic structural properties i.e. near ideal, large grain size and preferred orientation. This technique shows great promise for analytical inspection tasks requiring highly penetrating radiation such as security screening, medicine and non-destructive testing.

Synchrotron-based coherent scatter x-ray projection imaging using an array of monoenergetic pencil beams

Traditional projection x-ray imaging utilizes only the information from the primary photons. Low-angle coherent scatter images can be acquired simultaneous to the primary images and provide additional information. In medical applications scatter imaging can improve x-ray contrast or reduce dose using information that is currently discarded in radiological images to augment the transmitted radiation information. Other applications include non-destructive testing and security. A system at the Canadian Light Source synchrotron was configured which utilizes multiple pencil beams (up to five) to create both primary and coherent scatter projection images, simultaneously. The sample was scanned through the beams using an automated step-and-shoot setup. Pixels were acquired in a hexagonal lattice to maximize packing efficiency. The typical pitch was between 1.0 and 1.6 mm. A Maximum Likelihood-Expectation Maximization-based iterative method was used to disentangle the overlapping information from the flat panel digital x-ray detector. The pixel value of the coherent scatter image was generated by integrating the radial profile (scatter intensity versus scattering angle) over an angular range. Different angular ranges maximize the contrast between different materials of interest. A five-beam primary and scatter image set (which had a pixel beam time of 990 ms and total scan time of 56 min) of a porcine phantom is included. For comparison a single-beam coherent scatter image of the same phantom is included. The muscle-fat contrast was 0.10 ± 0.01 and 1.16 ± 0.03 for the five-beam primary and scatter images, respectively. The air kerma was measured free in air using aluminum oxide optically stimulated luminescent dosimeters. The total area-averaged air kerma for the scan was measured to be 7.2 ± 0.4 cGy although due to difficulties in small-beam dosimetry this number could be inaccurate.