Detection tasks in medical imaging

Example LUNA16: detection of pulmonary nodules in 3D chest CT images.

We discuss several traditional image detection methods.

Concepts:

  • Template matching
  • Matched filter
  • Hit criterion

Template matching

Simple but powerful. We've seen filters that show the highest response on themselves, meaning they respond to similar structures. This is essentially the idea of template matching. If you want to detect nodules attached to a vessel, design a bank of filters that matches a tube and an attached sphere from different perspectives.

This assumes we know the shapes to use. Instead, we can also learn the template.

Matched filter

A matched filter is one that is learned based on the appearance of the object we want to detect.

Example: stent strut detection in Intravascular Ultrasound images (IVUS). The IVUS is a rotating probe that makes ultrasounds within arteries.

This way you can detect areas that may rupture/collapse and that would require stenting.

Cardiologists then need to know that the stent is "well-opposed" meaning that it indeed counters the narrowing of plaque etc.

In the IVUS images, it was asked to cardiologists to click/label true positives and also potential false positives.

To get the matched filter, they than did a statistical analysis to get the average appearance of a true positive struct. Using this matched filter provided good results.

You can also use these training labels to learn which area size should be considered for true positives such that false positive are excluded. This is handy for post-processing out the false positives.

One could also use a prior to only select structs along the elliptical shape where struts can be expected.

Feature-based classification

See ISMI-2d-Classification.md .

Gabor filters, binary local pattern, etc. all combined into a long feature vector per pixel/sample. Then use a classifier on it, e.g. kNN or AdaBoost or Random Forests.

Detection performance metrics

When do we decide that we have "hit" the right target? We need a hit criterion.

If we know the common target area, then we can for example say we hit the target if the center of our automatic detected result has distance d from the target center such that $d < 2r$ with radius r. We use two r because often even nodular targets are more irregularly shaped than a perfect ellipsoid.

The particular hit-criterion depends on your application.

Because we do not care about background points, we typically only consider TP, FP, and FN.

Again, compute sensitivity, specificity, f1 etc. Sensitivity is also called the "detection rate" in this context.

FROC

For each detection we can provide a likelihood value, so an FROC curve is applicable.

ROC curve is usually used in binary classification, whereas a Free-ROC (FROC) curve is used in detection. What's the difference?

The FROC is very similar to ROC, but the difference is on the x-axis. In ROC the x-axis plots lists the false positive rate as 1-specificity. In FROC we instead have the average number of false positives (in terms of wrongly detected objects) per scan.

One major advantage of such a (F)ROC curve is that you can go to medical specialists and ask them: how many false positives per scan is acceptable? This is easier to understand than asking them what f1-score they want.