Database-assisted low-dose CT image restoration.

PURPOSE Acquiring data for CT at low radiation doses has become a pressing goal. Unfortunately, the reduced data quality adversely affects the quality of the reconstructions, impeding their readability. In previous work, the authors showed how a prior regular-dose scan of the same patient can efficiently be used to mitigate low-dose artifacts. However, since a prior is not always available, the authors now extend the authors' method to use a database of images of other patients. METHODS The authors' framework first matches the low-dose (target) scan with the images in the database and then selects a set of images that contain anatomical content similar to the target. These "priors" are then registered to the target and form the set of regular-dose priors for restoration via an extended nonlocal means (NLM) filtering framework. To accommodate the larger spatial variability of the patient scans, the authors subdivide the image area into blocks and perform the filtering locally. The database itself is first preprocessed to map each image from its 2D image space to a corresponding high-D image feature space. From this encoding a visual vocabulary is learned that assists in the query of the database. RESULTS The authors demonstrate the authors' framework via a lung scan example, for both streak artifacts (resulting from smaller projection sets) as well as noise artifacts (resulting from lower mA settings). The authors find that in the authors' particular example case three priors were sufficient to restore all features faithfully. The authors also observe that the authors' method is quite robust in that it generates good results even when the noise conditions significantly worsen (here by 20%). Finally, the authors find that the restoration quality is significantly better than with conventional NLM filtering. CONCLUSIONS The authors image restoration algorithm successfully restores images to high quality when the registration is well performed and also when the priors match the target well. When the priors do not contain sufficient information, the affected image regions can only be restored to the quality achieved with conventional regularization. Hence, a sufficiently rich database is a key for successful artifact mitigation with this approach. Finally, the blockwise scheme demonstrates the potential of using small patches of images to form the database.