[ RadSafe ] Statistics - evaluation of digital images

Bjorn Cedervall bcradsafers at hotmail.com
Mon Mar 12 08:05:20 CDT 2007

Dear RadSafers,

I recently ran into a question related to comparison of different X-ray 

Various reference images where given a zero (0) value and then different 
image processing techniques where used whereafter some radiologists were
asked about the quality (clinical value for diagnostic purposees) of the new
(processed) images - ranking the pictures by giving them a subjective value
on a five-stepscale: -2, -1, 0, +1, +2.

The outcome of a "typical" test for five radiologists could then look like
the following:

Radiologist          Score
1                        +1
2                          0           (image of about the same clinical 
value as the reference)
3                        +2           (much better image than the reference)
4                         -1           (image of less quality than the 
5                        +1

(Average: 3/5 = +0,60)

The reference is always zero so these five scores could be seen
as paired comparisons:
0, +1
0,   0
0, +2
0,  -1
0, +1

Reasonably a distribution like this has no tails as there is a lower and an 
upper bound.

Now, I want to test if various paired rankings of this kind are 
statistically different.
My question is what a distribution of this kind is called (contrasting 
gaussian and
other types of distributions) and what the appropriate statistical test 
would be
(some kind of paired t-test??).

I apologize for this question being somewhat outside the core of radiology 
it is related to efforts to lower diagnostic doses.

Any help will be greatly appreciated,

Bjorn Cedervall      bcradsafers at hotmail.com

With tax season right around the corner, make sure to follow these few 
simple tips. 

More information about the RadSafe mailing list