Segmented Attenuation Correction

Conventionally, 2D smoothing is applied to transmission sinograms before dividing into the blank scan to determine attenuation correction factors (ACF) [65, 93]. However, this method of processing has the undesirable effect of causing a mismatch between the reso lution of the emission and transmission data and is not completely effective in controlling noise propagation [66]. An alternative approach is to reconstruct attenuation coefficient images derived from the transmission and blank data and then segment the images into a small number of tissue types with a priori known attenuation coefficients. For example, Huang et al [94] described a method where the operator manually defines the body and lung outlines on the attenuation images. After assigning attenuation coefficients to these regions, noiseless attenuation correction factors are then calculated by forward projection. This method was shown to provide equivalent results to measured attenuation correction using a transmission scan acquired for approximately one quarter of the time.

This approach has been extended by several investigators by automating the determination of lung and body outlines using various morphological operators and heuristics. For example, Xu et al used a simple thresholding method to segment attenuation images into three discrete regions: air, lung and soft tissue [95]. Unsupervised image segmentation is more practical than the manual approach, particularly for large data sets such as those encountered in whole body PET.

The main problem with segmented attenuation correction is that there is a large degree of variability in tissue densities from patient to patient, particularly in the lungs. Assigning the same population average value to the lung regions of each patient may lead to significant bias. An alternative approach is to calculate the histogram of j values for each patient study and assign values based on an assumed probability distribution for the lung and soft tissue components of the histogram [66]. More recently, Bettinardi et al described an adaptive segmentation method based on a fuzzy clustering algorithm [96]. This method is also based on the histogram of j values but it automatically determines both the number of tissue classes that can be supported by the data (based on the variance in the images) and their centroids. The method is sufficiently general that it can be applied to any region of the body and is able to distinguish bony structures from soft tissue given adequate counting statistics (Fig. 5.20).

As segmentation is a non-linear, non-stationary process, the various sources of error discussed are difficult to predict. Therefore, not only is the accuracy of the method slightly inferior to measured attenuation correction, but it is also less predictable. This may be a serious drawback in studies where reliable estimates of quantitative values, including physiological variables, are required. However, given the improvement in SNR that segmentation provides compared with conven tional transmission processing, these disadvantages may not be important in many clinical applications.

Was this article helpful?

0 0

Post a comment