Archives

  • 2019-07
  • 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • 2021-03
  • of SVM classifier with RBF kernel is O d

    2020-08-18

    of SVM classifier with RBF kernel is O ¼ ðd nsvÞ where d is 856 829 classifier on single, two views and fusion-based scheme number of input dimensions and nsv is the number of support 857 830 assessed using average of the free-response ROC (FROC) vectors [49]. 858
    831 curves of the two subsets. The comparison as such shows
    832 slight improvement in the case-based sensitivity; but case- 6. Discussion 859
    833 based performance is more significant as far as reliable aid to
    834 the radiologists in clinical diagnosis is concerned. Area under
    835 ROC curve (AUROC) showing performance of classifier on TMC FFDM images used in our study cover larger breast area and the 860
    839 difference between single view and fusion-based scheme is limitations of the proposed methods lies with the detection 864
    840 statistically significant or not. Usually the difference denoted accuracy of nipple centre and lesion centre. As mentioned in 865
    841 by p value less than 0.05 indicates that results from clinical the dataset, the border and nipple of the selected images is 866
    Please cite this article in press as: Sapate S, et al. Breast cancer diagnosis using abnormalities on ipsilateral views of digital mammograms.
    Table 8 – Comparison of single view, two view and fusion-based scheme.
    Sensitivity Single view Two-view Fusion
    FPs/I FPs/I FPs/I
    867 clearly visible in the profile and hence detected successfully. generate a correspondence score, represent suspiciousness of 915 868 Actually, in some cases the nipple may not be in the profile or the candidate lesion on either of the ipsilateral views. 916 869 border sometimes is not detectable successfully. In all such Consequently, this correspondence score of a lesion helps in 917 870 cases the radial distance of the lesions from nipple centre may matching it KPT 330 with the most appropriate candidate on the other 918 871 be wrong. The width of KPT 330 annular region can also be incorrect view. Although there is no substantial improvement observed 919 872 leading to the failure of the proposed method. The accuracy in the results (sensitivity) in Tables 6 and 7 as compared to that 920 873 may come down with increasing number of such cases. of in Tables 4 and 5, the FPs/I is reduced to a satisfactory extent. 921 874 However, we are trying to modify our earlier algorithm to More number of possible pairs in the proposed method are the 922 875 correctly delineate the border and identify the nipple position bottleneck in the overall performance. Even though the lesion 923 876 correctly especially with difficult cases in our future work. on one view is not paired with any other lesion on other view, 924 877 Consistent and accurate positioning of the breast during the single view characterization can yield the classification 925 878 mammography is essential for our method to acquire accurate results as shown in Fig. 4. 926 879 distance metrics for correspondence score. Incorrect and The final stage in the proposed work for improving lesion 927 880 inconsistent positioning may lead to lower accuracy that detection is the fusion of two view correspondence scores 928 881 was obtained on the dataset in active transport study. with the one-view abnormality score. The fusion process is 929 882 The authors in Refs. [9,46] showed that there is a high simple and computationally simpler than approaches de- 930 883 correlation between distances from the nipple to the centroid scribed in the literature. The metric shown in terms of Kappa 931 884 of lesions on CC and MLO views. The proposed work has indicates the slight correlation between observed and 932 885 leveraged this idea to devise a mechanism to choose lesion- expected accuracy which can be improved further. However, 933 886 based parameters as specified through Eqs. (5)–(7) in Section better features can be selected to improve the correspondence 934 887 4.3.1 to determine the width of annular region which is score and subsequently the accuracy in matching the lesion 935 888 different for every individual lesion on the either views. The pairs. 936
    889 results shown in Tables 4 and 5 prove that our selection of
    890 annular region is appropriate and works well on mammo- 7. Conclusion 937
    891 grams of larger as well as smaller breast area. It is observed
    892 that in few cases the correspondence score of TP–FP pair is
    893 better than that of TP–TP pair. We are investigating the reasons The significantly vital concurrent analysis for estimating 938
    894 behind such exceptional cases. likelihood of cancer of the abnormal lesions on the ipsilateral 939
    895 The purpose of minimizing the false pairs of lesions using views of the mammogram is described in this article. The