Accuracy

Discussion on Accuracy

What is "accuracy" in a colony counting machine?
Today, most petri dishes generated in research are counted by trained humans. A colony counting machine, therefore, is judged by how well it emulates a well trained, well rested and motivated human counter.

So how does MACE emulate a human counter?
Humans can see very small objects, if the contrast is high enough. We can see stars in the sky although their physical size is for practical purposes, zero. We can see and recognize objects of very low contrast, if they are big enough. Think of trying to read boilerplate on the back of a typical consumer contract printed in gray on light gray. Imaging devices can see objects that are smaller and lighter than a human equipped with a typical low power magnifying glass can see. To emulate a human counter we must measure the "visibility" of every object in the petri dish and eliminate those that fall below a "visibility limit". Many separate things determine the visibility of an object.

Chief among these are size and contrast. However, object shape, edge sharpness, background characteristics all have an effect on determining visibility. When objects are clumped together and overlap into groups, shape and density variations provide clues as to how many objects there are in the group.

In MACE many factors are considered when determining accuracy.

The graph shows the results of comparing machine counts to human counts on a set of 50 plates, each containing from 25 to 250 colonies. Each dot compares the logarithm of the machine count to the logarithm of the human count. These results are almost identical to a similar test comparing two human counters.