Compression
Studies were conducted with trained and untrained surgical pathologists
to determine the degree to which images could be compressed without losing
image 'quality.' Quality was left to the definition of the viewer, but the
trained pathologist considered diagnostic criteria as well as 'appearance'
while the untrained reviewers were making judgments more on aesthetics.
The results from all groups was clear and consistent with that in the literature,
viz. signficant image compression does not result in loss of image recognition,
comprehenion and/or analysis. Details are given below the image.
Specimens were generated from the slide archives of the Pathology Department
of Robert Wood Johnson University Hospital, New Brunswick, NJ and were an
ensemble which had previously been diagnosed using traditional methods up
to five years earlier. traditional methods up to five years earlier. Specimens
were digitized with a three chip, high resolution SONY 970 MD color camera
interfaced with a Coreco Occulus acquisition board. The board generated
both 24-bit and 8-bit color images using proprietary algorithms.
Six staff pathologist reviewed the digitized cases as they were transmitted
across computer networks and/or phone lines to high resolution monitors
located in the surgical pathology suites at Robert Wood Johnson University
Hospital in New Brunswick or in Hamilton, NJ. The pathologists were asked
to render diagnoses for using the digitized images corresponding to cases
which they had interpreted up to five years earlier using traditional microscopic
methods (specimens on glass slides). Clinical diagnoses rendered using digital
images were consistent with those reported using conventional microscopy
in 97% of the cases studied.
The images corresponding to those cases which had been correctly diagnosed
fell into three categories: hematological, histological, and stained needle
biopsies extracted from liver. The images were compressed to ten different
levels using wavelett and standard JPEG technique.
Five observers participated in the experiments. Four were certified
pathologists and one was a four-year medical student. All had normal or
corrected-to-normal vision and viewed the images binocularly. Each participant
was seated 45.7 cm away from the screen of an Silicon Graphics Indigo2 Extreme
Workstation. A 512 x 480 image covered 14.6x13.3 cm on the screen. None
of the pathologists had any recollection of previously seeing the cases.
The effect of image compression was measured under two distinct perceptual
criteria, just noticeable difference (j.n.d.) and largest tolerable distortion
(l.t.d) using a random number generator to select compression levels and
display images while a two-alternative forced choice (2 AFC) experimental
paradigm and ascending and descending staircase statistics were computed.
The observer indicated his/her choice by clicking on one of two graphical
bars. They could establish their own pace by delaying their response. The
data from the "diagnostic accuracy" component of the experiment
showed that hematological images, histological images, and liver biopsies
used in this study could be compressed by factors of 35,20, and 25 respectively,
showed that hematological images, histological images, and liver biopsies
used in this study could be compressed by factors of 35,20, and 25 respectively,
without compromising their clinical usefulness.
In the "just noticeable difference" experiment the observers
tolerate lower compression ratios. It is important to note that for the
same observer the l.t.d. diagnostic threshold almost always exceeded that
for the j.n.d. The j.n.d. component of the experiment showed that hematological
images, histopathology images, and liver biopsies used in this study could
be compressed by factors of 22.9, 13.9, and 16.7 without decreasing subjective
image quality. The results of this study suggest that although there is
fairly large disparity among individual expert observers in their capacity
to perceive distortion due to compression there is relative good agreement
among observers when asked to indicate how much compression is tolerable
before the diagnostic essence of an image has been compromised. In addition,
thresholds were shown to be strongly dependent upon image category (hematopathology,
histopathology, liver biopsy) but only weakly dependent upon specific images.