Low Dose Dynamic CT Perfusion
Dynamic CT perfusion is one of the most comprehensive functional CT techniques. It consists of scanning over time the same body region slightly before, during and slightly after a short and compact contrast material bolus injection. The CT values of each voxel over time are collected in the so-called time-attenuation curves (TACs). Contrast material is highly attenuating, and the change of its concentration over time is directly proportional to the change of CT value over time for that specific voxel. These contrast (and hence blood) dynamics reflect important voxel-specific functional information. Some examples are: blood flow (BF), which represents how fast the blood is transferred from the arteries to the tissue capillaries of that specific voxel, blood volume (BV), which represents the percentage of volume occupied by capillaries in that specific voxel, and mean transit time (MTT), which represents the average time needed from one single blood molecule to pass through that specific voxel.
An increase of BF and BV might reflect microvasculature proliferation, and it is then a marker of active tumor. Decrease of BF and BV are normally associated with ischemia. The main clinical applications of dynamic CT perfusion are thus in the field of oncology and neurology, both for diagnosis improvement and therapy monitoring.
The acquisition times are normally quite fast (less than one minute) and functional values are obtained via different physical models grouped under the name of dilution theory. The major drawback of this technique is the higher dose compared to conventional CT scans, due to the continuous radiation exposure. Unfortunately dilution theory models normally consist of deconvolution and/or derivative steps, which make the quantitative results susceptible to noise. One focus of our work is to develop robust algorithms for noise reduction in dynamic CT perfusion datasets, thus allowing for lower dose examinations. Unfortunately in many cases (white matter in the brain, ischemic tissues, etc.) the signal to noise ratio is very low, and also frequency spectra of signal and noise might be very similar. For these reasons, approaches that purely consider TACs as one-dimensional signals are slowly being abandoned. TACs fitting approaches might not deal properly with the big variety of TACs shapes (very different for tumors, ischemia, vessels, leakage, etc.). Filters in image space, as well as iterative reconstructions approaches are the most promising at the moment, since they allow exploiting the high dimension of CT Perfusion datasets and their redundancy. All images acquired in different time points are correlated to each other, thus they can be vectorized and merged together into one big data matrix (Casorati matrix), and both spatial and temporal smoothness can be enforced. Treating the problem as an intrinsic spatio-temporal problem allows for different approaches, like high dimensional total variation regularization, low-rank regularization, or incorporating dilution theory model in the reconstruction/de-noise step directly.
Figure 1: Left: One single slice of a noisy dynamic CT perfusion dataset. Contrast material flowing in and out the vessels can be seen, but enhancement of tissue is masked by noise. Right: Same slice filtered with our newly-developed method.
© dkfz.de
Figure 2: Left: Blood flow map derived from the noisy dataset. Different structures and defects sizes are hardly recognizable. Right: Blood flow map derived from filtered dataset. A defect in the left hemisphere is clearly visible. Blood flow is decreased due to an occlusion in the left internal carotid.
© dkfz.de