We randomly selected 60 unenhanced chest LDCT studies (30 males,
30 females aged 55-74 years) (Figure 1).
as the algorithm would allow for detection of pre-clinical liver disorders in a large population.
Manual liver density was measured with >100mm2 region of interes (ROI) in liver segments and than averaged.
As liver density in various diseases varies from -40 HU to +100 HU [11],
we deemed ROI-based density segmentation inaccurate and created an algorithm that takes liver shape into account.
An algorithm for liver densitometry should also account for zones with different density (i.e.,
focal lesions or “geographic” steatosis) and quantify them separately.
The program has two stages.
First,
liver localization and segmentation.
The automatic segmentation algorithm consists of the following steps:
1.
Liver position approximation based on the adjacent skeletal tissue.
If this step does not detect liver,
the algorithm is terminated.
2.
Delineation of liver position and boundaries via a correlation algorithm using liver shape database.
The normal shape and size of the liver may vary [6].
However,
several types of liver shape can be distinguished based on the size of the left and right lobes.
Thus,
we identified six liver shape variants and manually created binary liver masks from axial slices.
Three-dimensional liver shapes for the database were generated automatically from binary masks (Figure 2).
3.
Boundary-based mean liver density quantification.
4.
Precise total liver volume segmentation using liver boundaries and its density (Figure 3).
Second,
measurement of all liver voxels (including vessels and bile ducts).
Our algorithm accounts for possible liver heterogeneity.
If the segmented volume has areas with different density (i.e.,
a focal lesion),
the algorithm will calculate HU values for each one.
The result is a histogram based on the average density values of detected liver borders (Figure 4).
It reflects the voxel-to-density ratio.
Liver lesions will be reflected as an extra peak on the histogram.
The cutoff value is 20% from average liver density.
We used a correlation method for matching three-dimensional model with patient’s liver.
The average liver density is then calculated based on the detected area.