Inconsistent read from `.mha` vector image

[tbirdso]

Description

Behavior is observed where vector pixel data is read inconsistently from a small .mha data file.

Steps to Reproduce

1. Download the data file from the ITKUltrasound data folder on Girder: rf_voltage_15_freq_0005000000_2017-5-31_12-36-44_ReferenceSpectrum_side_lines_03_fft1d_size_128.mha. The data file is a 3D image with a single voxel that is a 31-channel vector.
2. Repeatedly read in the 31-channel vector image in Python and access the 29th or 30th channel

>>> import itk
>>> pixel_type = itk.VariableLengthVector[itk.F]
>>> image = itk.imread(r'path/to/rf_voltage_15_freq_0005000000_2017-5-31_12-36-44_ReferenceSpectrum_side_lines_03_fft1d_size_128.mha',pixel_type=pixel_type)
>>> image.GetPixel(0).GetElement(29)
1.2316740878152266e-39
>>> image = itk.imread(r'path/to/rf_voltage_15_freq_0005000000_2017-5-31_12-36-44_ReferenceSpectrum_side_lines_03_fft1d_size_128.mha',pixel_type=pixel_type)
>>> image.GetPixel(0).GetElement(29)
1.6303663969778794e+20
>>> image = itk.imread(r'path/to/rf_voltage_15_freq_0005000000_2017-5-31_12-36-44_ReferenceSpectrum_side_lines_03_fft1d_size_128.mha',pixel_type=pixel_type)
>>> image.GetPixel(0).GetElement(29)
1.6303663969778794e+20
>>> image = itk.imread(r'path/to/rf_voltage_15_freq_0005000000_2017-5-31_12-36-44_ReferenceSpectrum_side_lines_03_fft1d_size_128.mha',pixel_type=pixel_type)
>>> image.GetPixel(0).GetElement(29)
1.2316740878152266e-39

Expected behavior

Pixel output value is consistent

Actual behavior

Channel 29 alternates between 1.2316740878152266e-39 and 1.6303663969778794e+20

Channel 30 alternates between 2.1019476964872256e-44 and 0.15648820996284485

Reproducibility

Alternates evenly, approximately 50%

Versions

ITK 5.3rc03

Environment

Windows, Python 3.8.5

Additional Information

[tbirdso]

Hi @aylward @jjomier, do either of you perhaps have thoughts on how best to investigate this issue with reading from a MetaImage file?

[jjomier]

@aylward let me know if I should investigate this. Looks like either the writing or reading of the file is incorrect (haven't tested C++)

[aylward]

Weird!

I'm interested in where this inconsistency may be introduced.

1. Try converting the file to .nii and see if the problem persists.
2. Try converting the image to an array (and/or an array view) and see if the problem persists
3. Try using an uncompressed metaimage...and see if the problem persists.
4. Read it into a different variable every time (instead of overwriting image every time)...and see if the problem persists

I think that covers most of the points in the pipeline that could be introducing this bad behavior.

Thanks for finding this!!!

s