dchansen / rabbitct Goto Github PK

What steps will reproduce the problem?
1. Use RabbitCTRunner to test implementation
2. Observe computed root mean square error

What is the expected output? What do you see instead?

The RMSE of Thumper for the 512^3 volume is 0.0207 HU.
This result has been computed with 32-bit float values.
The required HU scaling has been applied.
(FACTOR =  1.2076081037521362, OFFSET = 24.0)

When the volumes are rounded to short integer precision,
which is the storage format of the reference volumes,
the RMSE increases dramatically to 0.115 HU.

However, RabbitCTRunner displays a value of 0.157 HU.
This deviation is caused by the use of a discontinuous
transfer function for the HU scaling in the following line:

float HU_ax = rec_volume[i] > 0.0f ? US_ROUND( (HUS[0] * rec_volume[i] + 
HUS[1]) ) : 0.0f;

0.000f => 0.0f
0.001f => ~ 24.0f

This is also the reason why the maximum absolute error
is displayed as 24.0 HU by RabbitCTRunner.
The real maximum absolute error of Thumper is 0.662 HU.

What version of the product are you using? On what operating system?

The latest version of RabbitCTRunner.
The operating system is not relevant to this issue.

Please provide any additional information below.

The discontinuous transfer function can be fixed easily.
However, the computed RMSE is affected dramatically by
storing the reference volumes in short integer format.

What steps will reproduce the problem?
1. Checkout trunk under Linux
2. Run CMake, and make to compile
3. Comping fails due to an ambiguous function call

What is the expected output? What do you see instead?

I would expect the runner to compile without errors, but instead I see

/rabbitctrunner.cpp:321: error: no matching function for call to 
‘max(double&, const float&)’

What version of the product are you using? On what operating system?

SVN trunk on Ubuntu 10.04.1 LTS, 64-bit.

Find a patch attached that fixes the issue.

including <string> header in Visual C++ to use std::string
rabbitctrunner.h

#include <string> 

There is an error in the reference implementation of the backprojection 
algorithm in the source files "LolaBunny.cpp" and "LolaOpenMp.cpp". This error 
also affects the reference volumes provided in the data set files 
"rabbitct_512.rctd" and "rabbitct_1024.rctd".

In the function "inline double p_hat_n(double x, double y)",
which performs a bilinear interpolation at position (x, y),
input values in the interval ]-1, 0[ are treated incorrectly:

inline double p_hat_n(double x, double y)
{
    int i = (int)x;
    int j = (int)y;
    double alpha = x - (int)x;
    double beta = y - (int)y;

    ...
}

The integer cast truncates the floating point numbers towards zero,
while the desired functionality is to truncate towards minus infinity.

current: y = -0.5 => j = 0, beta = -0.5, (1 - beta) = 1.5
desired: y = -0.5 => j = -1, beta = 0.5, (1 - beta) = 0.5

In this example, the input coordinates used for interpolation
are 0 and 1, and the weights are -0.5 and 1.5, respectively.
Therefore, the fact that the border of the input images
is set to zero does not prevent the error.

In the geometric configuration of the RabbitCT data set,
the affected voxels lie at large values of the z coordinate.

A correct, but slower version, of the function p_hat_n 
could use the function floor() as follows:

inline double p_hat_n(double x, double y)
{
    int i = static_cast<int>(floor(x));
    int j = static_cast<int>(floor(y));
    double alpha = x - i;
    double beta  = y - j;

    ...
}

What version of the product are you using? On what operating system?

Version: RabbitCT-1.0.3-src.zip
The reported error is independent of the operating system.

Best regards,
Timo