UC San Diego

In computed tomography (CT) imaging, if metal is present in the scan, it gives rise to streaks and shadows called metal artifacts. We consider two applications of CT, radiology and luggage screening for aviation security. In radiology, metal artifacts make it difficult to evaluate anatomical structures. In luggage screening, computations on metal-artifact degraded images give rise to false alarms. Therefore metal artifact reduction (MAR) is an active area of research. For medical imaging, we improve upon a class of MAR algorithms that are often called sinogram completion methods. The sinogram (Radon transform) contains the log-attenuation measured by the scanner. In sinogram completion methods, portions of the sinogram contaminated by metal are replaced with estimates of the underlying data. Our improvement comes from segmenting artifacts from anatomy, based on their spatial and intensity distributions. Segmentation yields an intermediate image which when forward-projected, guides the sinogram completion. The corrected sinogram is reconstructed into the final image. We applied our algorithm to CT scans of the head and found that our results improved upon the state-of-the-art. In luggage screening, the variety of scanned articles is larger and the amount of metal is greater, therefore assumptions cannot be made on spatial and intensity distributions. Our strategy here is a hybrid one, combining numerical optimization with sinogram completion. The numerical optimization de-emphasizes metal-contaminated projections. We compared our method to previously published MAR algorithms qualitatively and quantitatively. Our method reduces metal artifacts and preserves more image details than the compared methods. We also developed methods to evaluate the accuracy of segmentation algorithms in CT. The first method is based on mutual information of machine segments (MS) against ground truth (GT) segments. Mutual information is computed from a confusion matrix that contains the quantity of a feature common to MS and GT labels. The second method is based on feature recovery. We compute optimal one-to-one correspondence between GT and MS labels, and extract total and systematic errors. The errors give us insights that can be used for improving the algorithms. The evaluation of these methods themselves was based on synthetic problems and human observer evaluation