[Insight-users] Measuring tumor diameter

Mon Dec 22 10:05:34 EST 2008

On Mon, Dec 22, 2008 at 1:59 AM, Karthik Krishnan
<karthik.krishnan at kitware.com> wrote:
> As Celina mentions, you need to look at clinical acceptability.
>
> The standards used at present are the WHO and the RECIST (Response
> evaluation criteria in solid tumors) criteria.
>
> - WHO takes the sum of the product of the longest diameters in perpendicular
> dimensions.
> - RECIST is simpler (and newer) and takes the sum of the longest diameter.
>
> Both are made by radiologists on orthogonal 2D slices.
>
> In a clinical trial, for which we've written software in the past, these
> measurements are made by typically 3 doctors independently and their
> measurements are adjudicated by another doctor.
>
> I'm not arguing the merit of your problem, since its far better than WHO
> (1979) and RECIST (2000) standards, but you need to factor in its chances of
> being accepted into clinical practice.
>

Karthik,

In fact, "my approach" is just a specific way to automate the
calculation of the tumor diameter, which, as you correctly pointed
out, is required by both WHO and RECIST standards. I do not dare to
call for standard changes :)

Andriy Fedorov

>
> Richard Beare wrote:
>>
>> So it sounds like the current working definition of tumor diameter
>> used by surgeons/medicos is the length of the longest straight line to
>> fit inside the tumour body, and that in practice this is usually
>> estimated using 2d slices.
>>
>> In the ideal world you'd have a dataset of tumor images with diameter
>> estimates made by a variety of medicos against which you could compare
>> the various computational metrics - perhaps you have this.
>>
>> There appears to be at least one major assumption/approximation that
>> medicos are willing to make with this sort of tumor - they are happy
>> with a single measure that basically assumes that the tumor is
>> approximately spherical.
>>
>> So it seems to me that you have a number of alternatives. The first is
>> direct implementation of the manual estimation method you described.
>> As you say this will require computing something for all voxels on the
>> surface of the tumour, which is several thousand points, but I don't
>> think that will be prohibitively time consuming, at least compared to
>> actually segmenting the tumor well. There are a number of tricks that
>> might help speed things up.
>>
>> Here are a few thoughts:
>>
>> Computing the distance between all pairs of surface voxels is
>> feasible, but determining whether the intersecting line goes outside
>> the surface might be much more complex. Therefore an estimate of the
>> lower bound of distances that you are interested in would let you
>> avoid determining whether or not short lines fall outside the tumor. A
>> couple of estimation methods spring to mind - 1) measure the tumor
>> volume and determine the diameter of the equivalent sphere - any line
>> less than some significant fraction of this isn't of interest 2) Find
>> peaks in the euclidean distance transform, values of which represent
>> the radius of the biggest sphere that will fit inside the tumor - any
>> line less that a significant fraction of twice this value isn't of
>> interest. Then build a queue of voxel pairs sorted by distance between
>> them and check for valid line segments, starting with the longest
>> first. Stop when you find one that lies entirely within the tumor.
>>
>> You could then compare performance of this sort of approach to the
>> other options that have been raised. At least this is a direct
>> translation of the current practice.
>>
>> On Mon, Dec 22, 2008 at 9:34 AM, Andriy Fedorov <fedorov at bwh.harvard.edu>
>> wrote:
>>
>>>
>>> On Sun, Dec 21, 2008 at 11:20 AM, Steve M. Robbins <steve at sumost.ca>
>>> wrote:
>>>
>>>>>
>>>>> The approach I am currently considering is this:
>>>>>
>>>>> 1) go through the axial slices, find the one with the largest area
>>>>> 2) extract that slice contour
>>>>>
>>>>
>>>> This makes it seem like you have reduced the problem to 2D only.  Is
>>>> that desirable, or is that done due to convention (e.g. radiologists
>>>> traditionally look at stacks of 2D slices)?
>>>>
>>>>
>>>
>>> Based on my discussion with a neurosurgeon, they usually look at 2D
>>> slices in AP/LR/IS directions, but also they may also look at oblique
>>> slices.
>>>
>>> To understand maximum diameter, as I understand it, imagine the
>>> contour of the tumor segmented in a 2d slice. Then take a set of
>>> segments connecting all possible combinations of the contour points.
>>> For each segment, subtract the subsegment which is outside tumor
>>> contour (this will happen only for concave shapes). The updated this
>>> way length of the longest segment will be the maximum diameter.
>>>
>>> I assume, one can take all combinations of surface points in 3d,
>>> instead of looking at a slice with the largest area, but this will be
>>> very time-consuming to walk along the segments connecting all possible
>>> pairs of points. I simplify the problem to develop an initial
>>> solution.
>>>
>>>
>>>>>
>>>>> 3) go through all possible combinations of the contour points, find
>>>>> the pair of most distant points, and take this as a diameter
>>>>>
>>>>
>>>> Here you are measuring using the normal Euclidean distance?  For
>>>> example, an "L" shape would have the two most distant points be the
>>>> end of the two legs and the diameter would join them to form a
>>>> right-angle triangle?
>>>>
>>>>
>>>
>>> An "L" shape will have the maximum diameter equal to the length of the
>>> longer leg. With the diameter you suggest most of it will belong to
>>> the outside of the shape.
>>>
>>>
>>>>>
>>>>> 4) follow the line between the points in the previous step, and
>>>>> subtract the parts of the line that are outside the contour (this is
>>>>> how the tumor measurements are actually taken). This may change the
>>>>> measured diameter.
>>>>>
>>>>
>>>> I don't understand "subtract the parts of the line outside the contour".
>>>> In my "L" shape example, almost the entire line is outside; does that
>>>> mean you would say its diameter is zero?
>>>>
>>>>
>>>
>>> The diameter defined this way will indeed be almost 0. This will not
>>> be the maximum diameter though.
>>>
>>>
>>>>>
>>>>> 5) repeat steps 3 and 4 until the maximum is found after taking into
>>>>> account diameter parts outside the countour
>>>>>
>>>>
>>>> Again, I'm not sure what this means.
>>>>
>>>> Naively, I might expect that you want to measure a longest distance
>>>> through the shape.  In the "L" example, this would be the sum of the
>>>> two legs.
>>>>
>>>
>>> The diameter by definition should be measured along a line, you cannot
>>> sum up two legs...
>>>
>>>
>>>>
>>>> I googled a bit but did not find any precise explanation of how to
>>>> measure a "tumour diameter".  Can you explain a bit?
>>>>
>>>>
>>>
>>> I haven't found a precise definition either. Based on my
>>> understanding, the measure is very subjective, and it is basically
>>> what I described. Of course, it is not easy to visually identify the
>>> largest diameter.
>>>
>>> I hope I was able to clarify the problem for you a bit. If you find a
>>> better definition, please post...
>>>
>>> Thanks
>>>
>>> Andriy Fedorov
>>>
>>>
>>>
>>>
>>>>
>>>> Thanks,
>>>> -Steve
>>>>
>>>>
>>>> -----BEGIN PGP SIGNATURE-----
>>>> Version: GnuPG v1.4.9 (GNU/Linux)
>>>>
>>>> iD8DBQFJTmzn0i2bPSHbMcURAv55AKCNOZkG6mTwfvrnlaIlc4JSBERHrQCeMa4g
>>>> yFH+p3QqTwZ79yiOQzf9F/E=
>>>> =BrcE
>>>> -----END PGP SIGNATURE-----
>>>>
>>>>
>>>>
>>>
>>> _______________________________________________
>>> Insight-users mailing list
>>> Insight-users at itk.org
>>> http://www.itk.org/mailman/listinfo/insight-users
>>>
>>>
>>
>> _______________________________________________
>> Insight-users mailing list
>> Insight-users at itk.org
>> http://www.itk.org/mailman/listinfo/insight-users
>>
>
> --
> Karthik Krishnan
> R & D Engineer,
> Kitware Inc,
> Ph:  518 371 3971 x119
> Fax: 518 371 3971
>
>