Learning center

Medical histograms: scope, percentiles and corrected exports

Declare the selected image domain, read corrected percentile values in numeric order and verify complete selected tables with historical recording limits.

Written guide to a bounded, corrected export workflow. Medical Image Processing histogram_analysis describes an image’s selected intensity values. The earlier reviewed app snapshot matches the locally accepted percentile-table candidate. Section 6 separately identifies the later native capture and this addition’s dated source inspection. The earlier CT film and medical-current-reference.json predate the saved-percentile correction; their historical evidence does not establish current export correctness or video acceptance.

1. Preserve the source and declare the analysis scope

The reviewed input is spleen_8.nii.gz from Medical Segmentation Decathlon Task09 Spleen. Its source dimensions are 512 × 512 × 31, or 8,126,464 voxels. The loader’s recorded array order is 31 × 512 × 512. Preserve the source hash, original spatial metadata and declared array selection; a different dimension order alone does not imply a different number of voxels or justify changing anatomical axes.

The reviewed native control defaults to Plane (2-D). For the corrected whole-volume example, explicitly select Volume (3-D). The saved receipt should report native 3-D volume and all 8,126,464 selected values. For a plane, retain its array axis and index; for a four-dimensional input, retain the leading index as well. A central image shown beside a volume result is a visual preview, not evidence that only that plane was analyzed.

Keep the analysis mask, ROI and region labels explicit. The reviewed run supplies no analysis mask and uses the complete selected volume; it is not a spleen-only measurement simply because the dataset is named Spleen. ROI and multi-region routes need real masks or integer labels, and comparison routes need a real second image with confirmed matching source grid. Missing prerequisites should produce unavailable outcomes, not invented regions or alignment.

2. Set the histogram and percentile contracts

The corrected recorded request selects histogram_moments and percentile_analysis, with 128 bins and Volume (3-D). It retains the default data-range histogram and requested percentiles 1,5,25,50,75,95,99. The reviewed parser requires 1–64 finite percentile levels between 0 and 100 in strictly increasing order; duplicates and reversed levels are refused. Histogram bins must be between 2 and 4,096.

Moments and percentiles use the selected intensity values, with any declared analysis mask. Percentiles use NumPy linear interpolation. Histogram bins describe a distribution plot; changing the bin count is not a way to change the definition of a percentile. Explicit histogram ranges must be recorded separately from source and mask selection. The app bounds the selected source to 33,554,432 voxels and requires at least two values in the analysis domain; a refused input is not a zero-valued result.

3. Read the units, population and result status

The corrected receipt labels intensity units: unspecified. Preserve that label. A CT filename, negative values or a plausible-looking histogram is not sufficient to relabel these exported numbers as Hounsfield units. Check the loader’s intensity transformations and metadata before comparing another input or applying a physical interpretation.

The saved moments include mean −428.51336694532824 and population standard deviation 485.81054320686843 across 8,126,464 values. These are descriptive measurements of the complete selected domain, including its background. Population variance and standard deviation use the population convention; do not substitute a sample estimate silently. The negative recorded mean also makes a coefficient-of-variation ratio difficult to interpret as a positive relative spread.

Both selected routes completed, but the receipt retains validated: false, publishable_as_validated_result: false and diagnostic_status: not validated for diagnosis. Successful execution and independent checks of selected numbers do not establish a diagnostic or segmentation claim. Threshold, fitted-model and comparison routes have distinct meanings and are not certified by these two routes.

4. Pair each percentile label with its corrected value

The earlier table conversion iterated a nested mapping’s keys and could export percentile labels instead of computed intensities. The corrected conversion preserves each named numeric value. In the saved table, values[5] is an intensity of −1024, not the percentile level 5.

Requested percentileSaved intensity (units unspecified)
1−1024
5−1024
25−1002
50−114
75−18
95135
99258

Use requested_percentiles to establish numeric order, then read the corresponding values[level] field. JSON object keys may be sorted as text: values[25] can appear before values[5]. That serialization order is not percentile order. The saved median is −114; the interquartile range derived from the saved values is −18 − (−1002) = 984.

The guide check reopens all seven corrected values and compares them with the same run’s published numerical metrics. The separate final numerical-agreement receipt checks those saved values plus sample count, mean, population standard deviation, median and interquartile range. This refresh does not recompute the entire CT or add new scientific validation.

5. Verify the complete selected tables and export limits

  1. Retain the exact input and request: app, two selected routes, Volume (3-D), bin count, percentile list and any mask. Confirm the source and selected-domain identities in the receipt.
  2. Open the saved result ZIP. This local server example has seven members: manifest, published results, receipt, two PNGs and two JSON tables. The percentile table is viewer/tables/001_Declared_Percentiles.json; identify it by the manifest’s title rather than assuming the same index in every run.
  3. Verify all seven percentile values, their labels and sample count, and read the complete moments table. These selected numeric tables preserve numbers without substituting rounded display text. The guide checks both the percentile-fix export and the later complete-table export by their distinct hashes.
  4. Read the package limits: app_exports and viewer arrays are empty here, while published histogram-array entries explicitly say array_not_embedded: true. Complete selected metrics do not mean every histogram array or voxel is included. This is not evidence of a reloadable scientific session.
  5. Compare member hashes and byte counts with the saved verification inventory, and open the PNGs and JSON outside the result viewer. The member-hash inventory is an external acceptance artifact, not a promise that this package’s native manifest supplies checksums. Preserve the raw input and metadata for any later rerun.

Worked examples describe particular inputs, methods and recorded results. Check current availability and input requirements before planning an evaluation.

← Back to the learning center