tumours This is nothing but identifying correct TP TP
tumours. This is nothing but identifying correct TP–TP pairs.
are paired with every other lesion from MLO view and
The 155 TP–TP pairs from CC to MLO are now divided into 4 sets
subsequently the correspondence score of the pair is calculat-
of 31 pairs for training and set of 31 for testing using five-fold
ed. The lesions on the MLO view which are lying outside the
cross validation using SVM classifier. Similarly, 163 TP–TP
annular regions defined by the radial distances of all the
pairs from MLO to CC are divided into 4 sets of 32 pairs for
lesions on CC view are discarded. Thus, there are limited
training and set of 35 for testing using five-fold cross validation
number of pairs from which one with highest correspondence
using SVM classifier. The correspondence score of the lesion
score is selected. This highest score actually belongs to the
on CC view extended by combined features set forms a new
most likely abnormal lesion from suspicious lesions on CC
feature vector. The same experiment for lesions on MLO view
view. In the second experiment, all suspicious lesions on MLO
is performed. The performances of a SVM classifier to
view are selected one by one and are paired with every other
distinguish the suspicious lesions on CC and MLO view are
lesion from annular region on CC view. Subsequently, the
depicted in Table 6.
correspondence score of the pair is calculated. All the possible
The average case-based lesions diagnosis sensitivity is
pairs of lesions from CC to MLO view are divided into 4 sets of
75.91% (from 110 cases) at the cost of 0.69 FPs/I on an average.
104 pairs for training and remaining set of 104 pairs for five-
The sensitivity here is calculated based on the number of TP–
fold cross validation using SVM classifier. SVM classified 155 of
TP pairs classified correctly either as benign mass or malignant
all the pairs as TP–TP and 365 as TP–FP. All the possible pairs of
tumour from actual 110 pairs. The specificity in this KPT330 case is not
lesions from MLO to CC view are divided into 4 sets of 112 pairs
meaningful as we have not selected true negatives (TN) for
Please cite this article in press as: Sapate S, et al. Breast cancer diagnosis using abnormalities on ipsilateral views of digital mammograms.
Table 7 – Performance of SVM classifier using fusion scheme on DDSM dataset.
True malignant False malignant True benign False benign Total pairs Sensitivity FPs/I Kappa
Fig. 8 – Performance of SVM classifier using (a) ROC plot and (b) FROC plot.
812 classification and hence not considered. Similarly, the results trials are statistically significant. In our experiment, the p 842 813 on DDSM dataset are depicted in Table 7. value is segments found less than 0.02. Hence the proposed fusion-based 843 814 The average case-based lesions diagnosis sensitivity is scheme is superior to the single view approach. 844 815 73.65% (from 110 cases) at the cost of 0.72 FPs/I on an average. Table 8 compares the results of classifier in terms of FPs/I 845 816 Sometimes the classifier performance metrics such as sensi- for single view, two view and fusion-based scheme at three 846 817 tivity, specificity and accuracy are not enough. Especially, the different adjusted sensitivity levels. 847
818 classifiers built and evaluated on multiple imbalanced data
819 sets can be compared more reliably through the Kappa 5.5. Computational complexity 848
820 statistics. It gives a better indicator of how the classifier
821 performed across all instances of different class distributions. The preprocessing using fuzzy c-means algorithms has the 849 822 In our experiment as shown in Tables 6 and 7, the values can complexity of O ¼ ði d n c2Þ where i is number of iterations, d is 850 823 be interpreted as indicators of correlation between observed number of dimensions (d = 2 in our case), n is number of pixels 851 824 and expected accuracy. in the given image and c is number of clusters. The 852 825 Fig. 8(a) depicts the performance of the SVM classifier on computational complexity of segmentation algorithm is O = 853 826 single, two view and fusion image-based CAD scheme using (n). The computational complexity of feature extraction with 854 827 ROC curves for the classifying malignant tumours from benign GLCM of a image matrix size M M is O = (M2). The complexity 855 828 masses. Fig. 8(b) shows the case-based performance of SVM