br from CC and other from MLO view
550
from CC and other from MLO view. Every lesion on CC view is
are easy to identify. In this regard, a set of features which are
560
551
paired with every other lesion lying within the annular region
invariant to the positioning and the SQ 109 are chosen.
561
552
of MLO of the same breast. Similarly, every lesion on MLO view
The features chosen based on an intuitive probability that they
562
553
is paired with every other lesion lying within the annular
would extract useful information are explained briefly as
563
554
region of CC of the same breast. To correlate the lesions on CC
below.
564
555
and MLO views, a few similarity features with highest rank as
Pixelwise correlation (SF1): This similarity measure is based on
565
556
explained in Ref. [46] are extracted from every possible pair of
566
557
lesions. As experienced by the radiologists involved in this
Pearson's correlation coefficient. Pixelwise correlation of a
567
558
study, the TP–TP pairs demonstrate dominant geometrical and
suspicious lesion Lcc on cc view with any suspicious lesion
568
559
textural features than that of TP–FP or FP–FP pairs and hence
Lmlo on MLO view mammogram is defined using Eq. (8).
569
Please cite this article in press as: Sapate S, et al. Breast cancer diagnosis using abnormalities on ipsilateral views of digital mammograms.
is depicted in Fig. 4 and the same is explained with the help of
Pi;jðLccði; jÞ
L2ccÞðLmloði; jÞ LmloÞ
two scenarios as below.
i;j
ðLmloði; jÞ Lmlo
Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiððÞ
P
P
5743
where Lcc(i,j) and Lmlo(i,j) are the pixel value at ith row jth
(1) A lesion on CC view is paired with a lesion on MLO view and
575
column of lesion template whereas
L
cc and
L
mlo are the mean
for this pair –
576
pixel values of the lesion template of the suspicious lesions
577
being evaluated, on CC and MLO view respectively.
ð
58079
Correlation standardized by median (SF2): It is a variation of
Pearson's correlation coefficient and is defined using Eq. (9).
of the features.
(ii) Radial distance-based feature RF is calculated as
i;jðLccði; jÞ LccÞðLmloði; jÞ LmloÞ
P
where Ry is calculated using Eq. (1).
i;jðLmloði; jÞ LmloÞ
(iii) Correspondence score of this pair is calculated as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiððÞ
P
P
where Lcc
and Lmlo are the median pixel values of the lesion
ðRFði; jÞ þ SCði; jÞÞ
586
template of the suspicious lesions being evaluated, on CC and
587
MLO view respectively.
Similarly,
correspondence
score
Pixelwise mutual information (SF3): It is assumed that the
5910
similar lesions contain a lot of intensity pixels which are
The pair with highest score is assumed to be a TP–
TP pair, rest all are false pairs.
redundant. The purpose of this information measure is to
maximize the redundant information between the partici-
(2) A lesion on CC view paired with no lesion on MLO view:
pating lesions Lcc(i,j) and Lmlo(i,j). The definition for the
the correspondence score is set to zero.
mutual information is given in Eq. (10).
Though the lesions from ipsilateral views are taken,
the aim here is to assess the suspiciousness of the
individual lesion on one (CC) view using its correlation
with lesion on the other (MLO) view as additional
pðLccði; jÞ; Lmloði; jÞÞ
log
information. Firstly, the highest correspondence score
ð ccð
Þ
mloð
ÞÞ
Lcc
ð
ð
Þ
ÞÞ
is assumed as belonging to a true positive lesion on the
XÞ
CC view. Secondly, the correspondence score for true
5998
where p denotes probability. The probabilities in Eq. (10) are
positive lesion on MLO view is calculated. The
600
calculated based on a joint histogram of intensities of pixels
correspondence score can act as an extended feature
601
levels of Lcc(i,j) taken along the x axis of the histogram and the
of that individual lesion. Hence, ingestive feeders is used in fusion of
602
pixel levels of Lmlo(i,j) are taken along the y axis.
single and two view detection scheme covered in the
next section.
6054