Rework the histogram max/min value autocompute flag in the `Thresholding` module

[jhlegarreta]

Description

Decide if the flag allowing to compute and set the histogram maximum and minimum values in the Thresholding module filters should be exposed.

The current design does not allow to set or control the range of interest of the histogram if the minimum and maximum values are not allowed to be computed automatically. Hence, when setting it to false, the thresholded images not be meaningful.

Impact Analysis

Allowing to set the histogram maximum and minimum values in the Thresholding module filter may involve a non-negligible amount of refactoring and work.

Although this would make the module more consistent with the presence of the m_AutoMinimumMaximum flag, leaving outside part of the histogram values translates this into a non-binary thresholding or classification problem. Hence, the thresholding modules may need additional refactoring to use ITK classes that are able to label accordingly images.

Expected behavior

Get meaningful, controllable results when the itk::HistogramThresholdImageFilter::m_AutoMinimumMaximum flag is set to false for the Thresholding filters.

One way to solve this would be able to set the itk::Statistics::ImageToHistogramFilter's m_Maximum and m_Minimum values manually, and be able to set such an appropriate instance to the Thresholding filter (possibly to their histogram calculator member). It should be automatically ensured that those maximum and minimum values are within the dynamic range of the input image.

This involves some sort of logic for the cases where m_AutoMinimumMaximum is set to false, or some new logic that would remove that flag to set those values by default by the histogram calculator (e.g. the m_AutoMinimumMaximum = true case) or set them to values specified by the user. And it may also involve modifying the itk::ImageToHistogramFilter API.

This would involve creating an itk::Statistics::ImageToHistogramFilter histogram generator instance, e.g. Look

if (!autoMinimumMaximum)
{
  using HistogramGeneratorType = itk::Statistics::ImageToHistogramFilter<InputImageType>;
  typename HistogramGeneratorType::Pointer histogramGenerator = HistogramGeneratorType::New();

  // Set the maximum and minimum values to desired values
  (...)

  // Set the histogram generator instance to the thresholding filter
  (...)
}

The itk::Statistics::ImageToHistogramFilter::InitializeOutputHistogram method and the classes involved need to be carefully looked at.

Actual behavior

The produced generated output is not meaningful because when the itk::HistogramThresholdImageFilter::m_AutoMinimumMaximum flag is set to false, the itk::Statistics::ImageToHistogramFilter does not initialize its m_Maximum and m_Minimum values to meaningful values, and thus the result of the thresholding process is not meaningful:

1. Run any of the tests in the Thresholding module where the autoMinimumMaximum flag is set to false,
   e.g. itkMaximumEntropyThresholdImageFilterTestNoAutoMinMax)
2. Check that produced output image is not meaningful (i.e. no anatomical structure can be distinguished).
3. Check that the expected threshold value is close to the lower end of the dynamic range of the input image pixels (it is only close due to the histogram's binning effect). In this case, the concerned thresholding filters return the same threshold value (provided that the input image is the same), e.g. -31744 for cthead1.png.

Versions

master

Additional Information

See this comment and the context in PR #511.

Related to issues #1737, #1738.

@richardbeare your thoughts and insight is highly appreciated.

[blowekamp]

The well defined behavior of the ImageToHistogramFilter needs to be taken into consideration and these automatic thresholding filters follow the same behavior.

In the ImageToHistogram class when the AutoMinimumMaximum flag is true the the range of the input is used. When false it defaults to the maximum and minimum of the input pixel range. However if the user sets the HistogramBinMinimumInput or HistogramMaximum input then the use values are used for the range of the histogram. Additionally, this class uses the advanced behavior of the range value being pipelined inputs.

Setting the AutoMinimumMaximum flag to false is useful when the input is uint8. It is sometimes with uint16 and 256 bins when the input has full 16-bit range, and can work with narrower ranges when more histogram bins are used.

The referenced test appear to use the cthead1.png file as input. The pixel type of this file is uint8 RGB, which appears to be read as scalar int16. Presumably the range is not scaled when read and converted into this type. For this situation setting the AutoMinimumMaximum flag false is not appropriate. I do not consider this behavior here a bug, but improper usage.

However, add the ability for a user to manually set the min/max of the histogram is a useful new feature, and it would make the behavior consistent with the ImageToHistograaFilter.

[jhlegarreta]

Thanks for the insightful comments Brad.

So to summarize what it can be done to address this, it looks like two things could address the issue:

• Changing/correcting the pixel types for the input image in the tests.
• May be using other more appropriate input images for the AutoMinimumMaximum false test case and writing new tests if necessary.

And once that is made clear in the code, and solved in the tests, then the ability to manually set the min/max values could be considered to be added as a new feature.

Hopefully making the tests have a proper usage of the flag would involve less development time.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. Thank you for your contributions.