Educational demo. Not a medical device. Not for diagnostic use.

The machine as a second reader

Unit 8 of 8

See which criteria the eight inputs touch, where the model and the expert disagree, and which criterion decides it.

The model reads eight numbers, all of them from nuclei: three sizes, one texture, and three shape values off the traced outline. That is the whole of what it sees. The panel below lists them against the criteria they touch, and beside them the criteria nothing here measures.

Its weakest link is not the mathematics, it is the scale. The coefficients were fitted on 569 patients measured on a 1990s camera at an unknown pixel scale, and the bridge from our 0.25 microns per pixel to theirs was fitted on four benign slides. Treat its odds as a second opinion from a reader who has seen only nuclei.

In the drills you commit first, then the call appears, then you are asked who is right and which criterion decides it. Only your first commit is scored. Agreement is tracked separately, on the Progress page.

What the outline measures

The eight inputs are the 1993 definitions, computed on our own 64 point outlines. The panel below pairs each one with the criterion it touches, and lists what no measurement on this site captures.

C5 Malignant

An unequivocal statement that the material is malignant, and the type of malignancy identified should be stated whenever possible.

Risk of malignancy
99 to 100 percent
Management
If the triple test is concordant, proceed to definitive management; if discrepant, CNB or excision biopsy is mandatory.

Source: Street WN, Wolberg WH, Mangasarian OL. Nuclear feature extraction for breast tumor diagnosis. 1993. Wisconsin Diagnostic Breast Cancer dataset, UCI Machine Learning Repository.

The worked example

Loading the case library.

What the machine sees

The eight inputs

worst radius
mean of the three largest nuclei's mean radius. Touches nuclear enlargement.
worst perimeter
mean of the three largest nuclei's perimeter. Touches nuclear enlargement.
worst area
mean of the three largest nuclei's area. Touches nuclear enlargement.
worst texture
standard deviation of grey values inside the three largest nuclei. Touches chromatin, in part.
worst smoothness
local variation in radius length, three largest nuclei. Touches pleomorphism of the nuclear margin.
worst concavity
severity of concave portions of the outline, three largest nuclei. Touches pleomorphism of the nuclear margin.
worst concave points
number of concave portions of the outline, three largest nuclei. Touches pleomorphism of the nuclear margin.
mean concave points
number of concave portions of the outline, every nucleus. Touches pleomorphism of the nuclear margin.

Not measured here:

  • cohesion and architecture (sheets, three-dimensional fragments, dispersal)
  • myoepithelial cells and bare bipolar nuclei
  • nucleoli
  • cellularity of the whole smear
  • the background (stroma, mucin, cyst fluid, pus, fat)
  • necrosis
  • mitoses

A reader who sees only the first column can be right about the size of the nuclei and wrong about the smear. That is why your call comes first and the model answers second.

The reliability diagram and the calibration table are on the explainer, under Model. They show how far the odds can be trusted, on the 569 patients the model was fitted on.

Start the session

About 20 trials, 5 to 8 minutes. You can stop at any point.