Boulder Laboratory for 3-Dimensional Electron Microscopy of Cells
MTTEACH(1) MTTEACH(1)
NAME
mtteach - develop criteria for feature recognition
SYNOPSIS
mtteach
DESCRIPTION
This program is used to develop criteria for distinguishing
microtubules (or other small features) from other image features.
First one prepares a file of teaching points. This file is built on
a series of areas and contains a point for each real feature in each
area. The first two points in each area are two diagonal corner
points of the area, NOT features; the rest of the points for that
area should all be inside these points. The program has modest
ability to check for errors in the teaching file. It is permissible
to add a point to the teaching point list that falls in ANY of the
zones defined earlier in the list, not necessarily in the last
defined zone.
Given the zones described by the areas with teaching points,
the program reads in the images contained in these zones (these
images would typically already be filtered, or correlated with the
average feature). The program then allows one to set up a two-step
procedure for identifying features. In the first step, the program
finds all of the points where the pixel value is more extreme than a
threshold value. One can adjust this threshold so as to find as
many of the true features as is feasible, at the expense of finding
many false features as well. In the second step, the program uses
discriminant analysis on the pixels surrounding the extreme peaks to
find a discriminant function (a function of those surropunding
pixels) that best distinguishes the peaks that are true features from
peaks that are not.
It is possible either to examine single pixels to look for peaks, or
to form a simple sum of pixels within a circular window around each
pixel position, and look for peaks in those "window sums". In
practice this seems to be useless, so the instructions below will
refer to examining pixels rather than "window sums".
When a prompt prints a number in [ ], that is a default value which
can be selected by entering "/"
ENTRIES to the program:
Image file name
Name of file of piece coordinates if image is a montage, otherwise
. Return
Name of teaching point file
0 use individual pixel values to look for peaks,
. or 1 to use sums of pixels within a circular window.
IF you select sums of pixels, then enter two lines:
. Radius of the window in pixels (which need not be an integer)
. 0 to center window on a pixel, 1 to offset it by (0.5,0.5)
Distance to search from a peak to eliminate nearby extreme points.
. When the program finds a pixel more extreme than the threshold
. (specified below), it will search up to this distance away for
. other points above threshold, and record only a single peak at the
. most extreme pixel.
Number of pixels at edges of zones to leave out of search for peaks.
. Typically one would enter the feature radius here.
At this point, one may repeatedly enter threshold values. The
threshold can be either an absolute value (if>1) or a relative
value (between 0 and 1). The program will search for pixels greater
or less than the threshold depending on whether an absolute
threshold is greater or less than the mean, or a relative one is
greater or less than 0.5. If one enters a relative threshold, then
within each separate zone, the program will look for a fixed
fraction of extreme pixels; i.e. a threshold of 0.99 or 0.01 will
select the brightest or dimmest 1% of pixels for further
consideration. For each threshold value, the program outputs the
number of teaching points that have been found and the number of
extreme points (after contiguity elimination) that do not correspond
to teaching points.
Instead of selecting a threshold, one may enter:
. -3 to output points above threshold to a file - useful only for
. diagnostic purposes (one needs to specify a threshold again)
. -2 to loop back and respecify whether to do window sums, the
. search radius for eliminating contiguous peaks, and the edges
. -1 to do discriminant analysis, if there are fewer than 1000 peaks
When discriminant analysis is selected, if there are any true
features that are not peaks above threshold, then the program will ask
whether to include those points as true features in the analysis.
Enter 0 to exclude them or 1 to include them.
Next specify a division of a circular region into a series of
sectored rings, by entering:
Radius of the entire circular region
Radius of central circle, which is not divided into sectors
Annular thickness (outer minus inner radius) of the innermost
. ring, the one outside the central, undivided circle
Number of concentric rings (excluding central circle)
Number of sectors to divide rings into
0 to center the rings on a pixel, 1 to offset by (0.5,0.5)
The thickness of the rings will change (typically increase)
progressively from the specified thickness of the innermost ring so
as to fit the given number of rings into the desired total circular
area.
The program makes a variable out of the sum of pixels in each sector,
then finds a discriminat function, a linear combination of these
variables, that best distinguishes the true peaks from the false
ones. It then finds a criterion discriminant score that minimizes
the total number of extreme points that would be misclassified.
The choices after a discriminant analysis are:
1 to scan a range of discriminant scores, finding the criterion in
. that range that minimizes misclassified peaks. Enter a lower and
. upper criterion score and the number of scores to try in the range
2 to redo discriminant analysis after respecifying sectored rings
3 to store analysis parameters and solutions in a file for use by
. MTDETECT (the last "best criterion score" found by scanning is the
. value that will be stored in the file)
4 to loop back to setting thresholds for peak selection
5 to store the discriminant function in an image file as a
. coefficient for each pixel
6 to exit
HISTORY
Written by David Mastronarde, May 1989