1. DISCLAIMER.

ABSTRACT.
1.  The first and most difficult step in image analysis in anatomic
 pathology is scene segmentation, where each object of interest
 (usually, a nucleus) is delineated for measurement.
2.  In typical image analysis software, the user has a sliding
 grayvalue stick, which may be adjusted until the nucleus appears
 dense and the cytoplasm and background is whited out.
3.  Subjective variations in threshold setting may result in
 differences of size, shape, and texture of the nuclear edge,
 and resulting measurements.
4.  ISAP: Image Segmentation and Analysis in Pathology, is a public-
 domain image software application, which offers both manual and
 automatic thresholding methods.
5.  Automated thresholding in this system obtains repeatable results
 on routinely-stained cytology smears.


INTRODUCTION.
1.  In image analysis software, the user manually adjusts a sliding
 grayvalue stick, until the nucleus appears dense and the cytoplasm
 and background is whited out.
2.  Estimation of the nuclear edge depends upon this manual threshold,
 which is not necessarily repeatable from observer to observer,
 or even within the same observer.
3.  Subjective variations in nuclear edge determination may, in turn,
 result in differences of estimated size, shape, and texture of the
 nuclear edge.
4.  We propose an automated thresholding method, which removes some
 of the subjectivity in nuclear edge determination.


MATERIALS AND METHODS.
1.  ISAP: Image Segmentation and Analysis in Pathology, public-domain
 image software application, with source code written entirely in
 Microsoft (R) Visual Basic for windows, version 3.
2.  Accepts uncompressed black-and-white images in Targa (.TGA)
 or Windows (.BMP) file formats.
3.  Operates on IBM PC or compatible computers, with 5 Mb ram memory,
 2 Mb hard disk available, DOS version at least 5.0, Microsoft (R)
 Windows, version at least 3.1, and a 256 color display monitor.
4.  Results are stored consecutively in comma-delimited import format
 for popular databases (Quattro, Excel, Lotus), for further analysis.


MATERIALS AND METHODS.
1.  Papanicolaou-stained cytology slides from the vagina and uterine
 cervix, obtained in routine patient care setting at the Baltimore
 Veterans Affairs Medical Center (BVAMC).
2.  In each of 6 patients, 5 cells apiece, and 10 measurements
 per cell, for a total of 300 separate analyses.
3.  Black-white images obtained as uncompressed 756 x 486 = 367,416
 pixels, with 256 grayvalues per pixel.  40x microscope objective;
 CCD camera; ATVista image grabber board.


AUTOMATED THRESHOLDING ALGORITHM.
 1.  User makes a tracing to delineate a workspace, which includes
 the nucleus and a surrounding rim of cytoplasm and background.
 2.  Software makes a histogram of all pixels in the workspace.
 3.  At each division point between two consecutive bars of the
 histogram, software calculates the pooled standard deviation
 for the Student t test.
 4.  Pooled standard deviation for Student t test is minimized when
 there is optimal separation between the dark (nucleus) segment of
 the histogram and the light (cytoplasm) segment of the histogram.
                        
 INITIAL EDGE DETERMINATION.
 1.  All pixels with grayvalues below-threshold (dark) are initially
 assigned inside the nucleus;  all grayvalues above-threshold (light)
 are initially assigned outside the nucleus.
 2.  This initial method is typically flawed by small, dark artefacts
 outside the nucleus, and by large, light artefacts inside
 the nucleus.
 3.  The distinction of artefact versus true feature
 is somewhat arbitrary.
 4.  A light pixel surrounded at 3 of 4 edges or at 5 of 8
 corners-or-edges by dark pixels is changed into a dark pixel.
5.  A dark pixel surrounded at 3 of 4 edges or at 5 of 8
 corners-or-edges by light pixels is changed into a light pixel.
                          
 FINAL EDGE DETERMINATION.
1.  The intuitively simple act of tracing the edge of a closed object
 (e.g., nucleus) is surprisingly difficult in anatomic pathology
 images.
2.  The center-of-gravity of all dark pixels is assigned
 as an inside point.
3.  For each inside point, every dark pixel which has a shared edge
 or shared corner with that inside point, likewise becomes an inside
 point.
4.  Step 3 is repeated to exhaustion.
5.  All non-inside points become outside points.
                         
 FIGURE 1.  Bimodal histogram, divided into left and right segments
 along the abscissa at x=10.  The left arithmetic mean is 5;  the
 right arithmetic mean is 11;  and the Student t pooled variance is
 1.41.  For a good division of two modes of the histogram, the pooled
 variance value is minimized.
                      
FIGURE 2.  Bimodal histogram, divided into left and right segments
 along the abscissa at x=6.  The left arithmetic mean is 4.3;
 the right arithmetic mean is 9.8;  and the Student t pooled variance
 is 3.82.  For a poor division of two modes of the histogram,
 the pooled variance value is higher.
                       
FIGURE 3.  Image from a Papanicolaou-stained cytology slide of the
 vagina and uterine cervix.  a single cell nucleus has been outlined
 by the user, and a histogram of grayvalues is shown for all pixels
 in the selected workspace.
                      
FIGURE 4.  The pixel histogram from Figure 3 is divided between each
 two consecutive bars, and the pooled standard deviation for the
 Student t test is minimized when there is optimal separation between
 the dark (nucleus) segment of the histogram and the light (cytoplasm)
 segment of the histogram.
                 
FIGURE 5.  In the pixel histogram from Figure 3, repetitive selections
 of slightly different workspaces result in nearly the same automated
 threshold value each time.
                     
RESULTS.
 Average coefficient of variation for nuclear area for six patients
 (5 cells apiece, 10 measurements per cell):  0.8%, 0.4%, 0.2%, 0.7%,
 1.6%, 1.4%.
                      
CONCLUSIONS.
 1.  Subjective variations in threshold setting may result
 in differences of size, shape, and texture of the nuclear edge.
 2.  ISAP software offers both manual and automatic thresholding
 methods.
 3.  Automated thresholding by the minimum pooled variance method
 obtains repeatable results on routinely-stained cytology smears.