CTP515 Low Contrast Module
The low contrast targets have the following diameters and contrasts:
Supra-slice target diameters Subslice target diameters
2.0 mm 3.0 mm
3.0 mm 5.0 mm
4.0 mm 7.0 mm
5.0 mm 9.0mm
Nominal target contrast levels
Low Contrast Measurements
Low contrast performance in CT imaging is a difficult subject. Clearly it is related to the ability of a "system" to distinguish between two objects or regions with similar CT number or X-ray attenuation. It is known to depend on statistical noise levels, contrast of the signal, and the size of the signal. In many cases, small signals (in the 3-7 mm diameter range) with contrast levels of some 3-5 HU are typically difficult to discern even in relatively uniform backgrounds. However, it is well reported that the contrast level (C) and the diameter of the target (D), when combined as a product approaches a constant (C*D≅constant). Thus, a plot can be made of contrasts (C) that might be found at a given diameter (D).
Because of the difficulty of the subject, much thought went into the construction of the Image Owl low contrast algorithm. The original method, used during the beta release of the service was based on the Rose Model. This calculates a C-D diagram based on contrast, diameter, and noise levels of the scan. Specifically, from this model, a target was considered "detectable" when the CT# contrast is considerably greater than the noise (2-5 times). However, working with David Goodenough, Ph.D., a new method was developed.
The previous method calculated low contrast values of measured contrast and pixel noise based on a contrast to noise ratio greater than 2. Simply, using known target locations, a target was considered detectable if the CT# contrast was 2 times the noise. The method the service now employs, estimates noise from the standard deviation of the mean HUs of two rows of circles of each target size in the background.
In the current method for low contrast calculation, at each target size (2-9mm and 15mm), an inner and outer row of ROIs (circles) are generated in the background as shown above. The mean HU for each circle is calculated and the standard deviation (SD) calculated from the set of means. If the standard deviation from the outer and inner circles deviates more than 10%, the algorithm uses the set with the lower standard deviation. A detection level for each target size is calculated, with a perceptibility factor of 4, as 4*SD.
The Rose Model represents probably the simplest model for detection of circular targets in a relatively uniform background (other than the statistical noise fluctuations seen in medical images). Strictly speaking, the model applies to "white" noise, meaning the noise excursions are uncorrelated from point to point (pixel to pixel). This is not necessarily true in CT, but nevertheless works reasonably well.
In the Rose Model, the eye (or the computer in this case) puts a circle (ROI of Region of Interest) over a target with the same diameter as the target and estimates the mean signal intensity of the target. This is a visual or numerical integration of the CT numbers in the target as averaged over the target. This value is then compared to mean intensity values in identical sized circles in the background region where only noise is present. A test of significance is made to determine how many times the mean signal intensity would exceed the expected fluctuation in the regions where only noise is present. The best case usually requires the mean signal intensity to be at least two times the expected noise deviation in white noise, and probably twice that for CT correlated noise.
Consider the case of Signal (S) and Background Level (B); with CT# of signal, CTS; and CT# of background, CTB.
For CT, contrast is CTS - CTB; and the CT noise is σ(H). Then signal contrast to noise ratio is CTS - CTB/σ(H).
To estimate noise excursions in signal sized targets, we note that for white noise, the relative noise standard deviation decreases by the square root of the number of pixels in the profile. If the single pixel noise standard deviation is "σ " then the profile (area) noise is approximately σ (H)/√N; where N is the number of single pixels needed to obtain the full area of the circular target.
Consider a profile through a 5mm, 0.3% target in an ideal scenario. Here we estimate background as 50H; ∴0.3% target is 53H. Remember in CT, 1% change from water represents 10H.
Now consider the same profile, but with single pixel noise at approximately 2H present in the profile.
If a typical pixel dimension is ≈0.5mm, then in a 5mm diameter target we may have π(52)/4 mm2 or ≈20mm2 area. This is about 80 pixels, or 20/0.52. If then the single pixel noise is 2H is approximately 2H/√80, or 0.2H , for the area standard deviation noise. In this case, a 5mm target at 0.3% contrast should be seen quite well because the ration is approximately 15:1. This is well above the 2:5 ratio suggestion for detection with white noise and reasonably above the 4:10 ratio for correlated CT noise, which behaves as if the white noise level is approximately doubled.
In a more normal (regular dose scan) technique with a typical 20cm phantom, where the single pixel noise is more like 4H, the ratio now becomes approximately 7.5:1, getting closer to the 4:10 limits of detection suggested for correlated CT noise. Moreover, a 3mm target with 0.3% contrast (a signal area about 1/3 that of the 5mm target with N ≈26) would have 0.7H effective noise for the area standard deviation and contrast to noise ratio of only about 4:1. This would not reasonably pass the 4:10 detection ratio, and therefore not be considered detectable.
- D.J. Goodenough, Psychophsical perception of computed tomography images. Radiology of the Skull and Brain; Technical Aspects of Computed Tomography, T.H. Newton and D. G. Potts, Eds, Vol. 5, pp. 3393-4021. Mosby, St Louis (1981).
- D.J. Goodenough and K.E. Weaver, Factors related to low contrast resolution in CT scanners, Computerized Radiol. 8, Vol. 5, 297-308 (1984).