IEEE Consumer Electronics Magazine - September/October 2020 - 31

Figure 2. Example of manually segmented
mammogram "mdb006" from the mini-MIAS
database.

Figure 3. Corner coordinates and the rotation angle
', which needs to be compensated for.

images manually requires that all grayscale values are visible in the prints. Another requirement
is making sure that a segmentation line can be
easily and precisely detected afterward. Mammograms usually have dark background and relatively bright tissue, so we inverted images before
printing them on a white paper using a standard
monochrome laser printer. Background of
images now becomes white and a thin border
around each image locates the edge of the image.
Printed images, of 16Â16 cm, were handed to an
experienced radiologist for segmentation using a
dark marker for the breast-skin line, and a white
marker for the pectoral muscle boundary.
Figure 2 shows result of segmentation of the
mammogram "mdb006." The mentioned border
of each image is used to correct any misalignment caused by printer- and scanner-feeders.
The applied correction on the scanned
images was image rotation for the angle approximated as

n1 À n 2 þ n 4 À n 3
' ¼ arctan
(1)
m1 À m2 þ m4 À m3
where mi and nj are x and y coordinates of the
detected corners starting from the top right in
counter clockwise direction, as shown in Figure 3.
A perfect match to the original image was
achieved by applying image resizing in horizontal
and vertical direction by two scaling factors
sh;v ¼

Horig ; Vorig
Hscan ; Vscan

September/October 2020

(2)

where sh and sv are scaling factors used for
image resizing and H and V are horizontal and
vertical dimension in pixels, respectively.

Tablet-Based Segmentation
A modern, but more challenging, approach is
tablet-based segmentation. Nowadays, consumer
tablets meant for personal-use, with highly accurate capacitive touchscreens, are a standard and
inexpensive equipment. Mostly used are IPS-type
LCD screens with good viewing angles, but somewhat limited contrast. Using a precise capacitive
stylus pen, it is possible to draw segmentation
contours, but display parameters need to be
repeatedly checked so that all gray levels are visible. In the case of MIAS mammograms, that task
is not very difficult because of 8-bit images. In the
case of mammograms stored as DICOMs,
intended to be viewed on 10-bit grayscale displays made for diagnostic purposes, usage of
commercial equipment would be more difficult.
To extract segmentation masks of the dense tissue, we decided to use tablet-based segmentation. Mammograms suffer from nonconsistent
pixel intensities for the corresponding regions in
the image. This makes automatic segmentation
methods complex and requires existence of a
ground truth dataset to compare the quality of
automatically segmented regions. Masks are
drawn as a closed contour around the breast tissue considered to be dense, see Figure. 4. Applying flood fill operation, while making sure that the

IEEE Consumer Electronics Magazine - September/October 2020

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - September/October 2020