br However currently radiologists use single view
113
However, currently radiologists use single view CAD
article.
114
51 systems as a second reader for breast cancer diagnosis using
52
FFDM images [10]. These CAD systems have high false positive
2.
Literature survey
53
(FP) and false negative (FN) rates because of varying subtle
54
nature of abnormalities and limitations of imaging system
55
[5,11]. The single view CAD systems for mass detection have
To overcome the limitations of prompting single view CAD
116
56
good diagnostic accuracy but less positive predictive value;
systems, many researchers have attempted to develop multi-
117
57
and hence are not as accurate and reliable as like the
view-based CAD systems. The clinical studies on breast cancer
118
58
experienced radiologist [12].
have indicated that two view mammograms help in achieving
119
59
Some findings reveal that the sensitivity of digital mam-
more diagnostic accuracy than with one view [18]. In their
120
60
mography to detect breast lesions obscured by surrounding
pioneer research, Good et al. [19] demonstrated that the
121
61
and overlapping dense parenchymal GW311616 is as low as 30–
multiview features of a single physical lesion on ipsilateral
122
63
breast lesions during screening using mammography [14]. The
0.03 on ROC curves. This approach has limitations on
124
64
diagnosis of these nonpalpable lesions is confirmed after the
number of features and number of lesion pairs. Chang et al.
125
65
histopathological reports (HPRs) of the sample tissues
[20] investigated arc method with
Az = 0.79 and Cartesian
126
66
obtained using surgical or needle biopsy. However, the
straight-line method with Az = 0.78
to localize and match
127
67
patients as well as radiologists wish to avoid biopsy of benign
breast lesions on 571 CC views with lesions on 571 MLO views.
128
68
mass but without missing an opportunity to detect the
In this scheme, the measurements related to nipple location
129
69
malignant tumour. Thus, accurate detection and characteri-
and chest wall orientation were not automatic. Paquerault
130
70
zation of the suspicious lesions can help to minimize the
et al. [21] developed two-view detection technique and
131
71
degree of cosmetic disfigurement of the breast during
reduced FP rate from 2.1 FPs/I to 1.2 FPs/I with sensitivity of
132
72
treatment. In short, improvement in the detection perfor-
80%. Sahiner et al. [22] devised a new method combining
133
Please cite predatory release article in press as: Sapate S, et al. Breast cancer diagnosis using abnormalities on ipsilateral views of digital mammograms. Biocybern Biomed Eng (2019), https://doi.org/10.1016/j.bbe.2019.04.008
134
results of single view and two views analysis to reduce the
3.
Clinical dataset
136
for the overall detection of breast cancer. The only drawback
137
of these approaches is that a true mass may be missed when
The clinical data for this study is taken from Tata Memorial
195
138
the detected lesion is on either view and not on both views.
Centre (TMC), Mumbai, India. Institutional Research Ethics
196
139
Paquerault et al. [9] incorporated the geometric locations,
Committee of TMC, Mumbai has approved the usage of all the
197
140
textural and morphological features to correlate the suspi-
biopsy-proven mammograms of breast cancer patients. The
198
141
cious lesions detected on ipsilateral views for improving
performance of the proposed two views CAD scheme is tested
199
142
CAD performance. They improved the case-based detection
on 110 pairs of CC and MLO views of FFDM obtained from 110
200
144
information. However, the annular search region with 80
identity of all the patients prior to start the work. All the
202
145
pixel radial width is a problem for smaller breasts. The
images selected for this study have proper visible border and
203
146
nipple detection is manual and not automatic as required
the nipple is also in the breast profile. FFDM images cover
204
147
for a CAD. Engeland et al. [23] presented a method to find
patients with different densities of breasts belonging to all
205
148
corresponding suspicious lesions on CC and MLO view with
classes with respect to Breast Imaging Reporting and data
206
150
one correspondence link of TP lesion with FP lesion instead
(176 lesions) are malignant tumours and 22 (44 lesions) are
208
151
of actual TP lesion. A new multiview-based CAD system