Computed Tomography (CT) technologies typically use x-rays or microwaves to image the inside of a sample or object. National security applications for CT include stockpile stewardship activities and cargo inspection at airports. The primary challenge with CT is to find objects, materials, or other features of interest in images with limited resolution, clutter, artifact corruption, and/or errors introduced as part of the image processing.
LLNL computer scientists are developing a new processing pipeline that tightly integrates traditionally separated image-processing stages, including segmentation and reconstruction, for improved threat identification. They are collaborating with partners in academia, government, and industry to develop and implement novel CT approaches that reliably detect a broad range of features, such as physical threats in luggage, flaws in engineering components, or signs of aging in the nuclear stockpile.
Starting in 2012, LLNL computer scientists began reworking the existing image-processing pipeline—acquisition, reconstruction, segmentation, and threat detection—into an integrated and iterative process to produce more precise imaging results.
During reconstruction, algorithms process the projections from CT systems and correct for imperfections in the image quality. Image information taken in the acquisition phase is typically ignored after reconstruction, but preliminary assessments have shown that this “discarded” data can provide important confidence information that is useful during segmentation.
For example, the presence of metal produces major artifacts in CT images because it strongly attenuates the x-rays and thus can shadow portions of a container’s interior. For each image pixel, the confidence is computed from the accumulated attenuation of all x-rays contributing information to the pixel; the darker the shadow or streak, the less confidence information is available.
Artifacts in the form of streaks are difficult to compensate for in a stand-alone segmentation algorithm and can easily cause artificial merging or splitting of objects. LLNL’s new processing technique uses confidence information to derive statistical measures of the likelihood that two neighboring pixels belong to the same object. The confidence information, coupled with each individual measurement, will allow the system to assess whether the corresponding differences in densities are significant.
As a result, segmentations will naturally integrate the uncertainty into the measurements and the reconstruction by operating on the confidence values, rather than the original image. By adapting this approach to the high dynamic range of the data, computational experts hope to make all scanned objects, including bags, appear more similar, and therefore easier to process with a general segmentation algorithm.
LLNL is currently working with various partners to integrate these concepts into a working system, evaluate their effectiveness in practice, and prepare a plan for quickly transitioning them into the field. In an airport setting, for instance, these image processing improvements should allow explosives detection technologies to more accurately differentiate threats from nonthreats and thus reduce false-alarm rates and increase the system’s operational efficiency.