Archives

  • 2019-10
  • 2019-11
  • 2020-07
  • 2020-08
  • 2021-03
  • Difference in entropy SF Shannon s entropy is a measure

    2020-08-18

    Difference in SQ109 (SF4): Shannon's entropy is a measure of 4.4. Fusion of single and two view information
    the average information carried in a pattern, which is widely
    used to quantify the smoothness of image texture. Entropy
    usually is relatively higher in heterogeneous regions and is As illustrated in Fig. 4, the radiomic features of suspicious
    low in homogeneous patterns. The absolute difference lesions are fused with the two views correspondence score to
    between the entropy of the regions on ipsilateral views is generate the final feature vector. For the single view CAD
    considered as a measure of dissimilarity. It is presented in system, a total of 72 textural features and 7 geometric features
    are extracted from each of suspicious lesions. The maximum 4
    lesions per image are considered for the sake of simplicity.
    This limit allows us to consider lesions on both views for
    pairing while keeping the FP rate as low as possible. The
    n
    Lccði; jÞ
    selected suspicious lesions are allowed to undergo two views r> ð
    ccð
    mlo ð
    Þ
    analysis as illustrated in Fig. 4. As described in Section 4.3.2, a
    X
    correspondence score is calculated using radial distance-
    based features and similarity features of lesions on CC and
    4.3.3. Computing correspondence score
    MLO views. This correspondence score is extended with single
    618 The location-based information and the similarity features of view-based feature vector of 79 features. In order to improve
    619 the corresponding pairs of suspicious lesions from ipsilateral the overall detection performance, malignancy score of a
    620 views are used to compute the correspondence score for all single view CAD system is combined with correspondence
    621 lesions. The features are assigned with the weight factors score of two view CAD system. The detected lesions are
    622 based on their rank as explained in Ref. [47]. The procedure classified as either malignant tumour or benign mass using
    623 followed to compute the correspondence score of a lesion pair SVM classifier. To avoid the biasing in the SVM classifier and to
    Please cite this 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 
    Table 1 – Results after segmentation on TMC dataset.
    True disease status w.r.t. HPR Results by proposed segmentation method Sensitivity FPs/I
    Abnormal Normal Total
    Table 2 – Results after segmentation on DDSM dataset.
    True disease status w.r.t. HPR Results by proposed segmentation method Sensitivity FPs/I
    Abnormal Normal Total
    684 test its performance, 5 random partitions in the training to segmentation for detecting suspicious lesions on images of 712
    685 testing ratio of 4:1 are employed [48]. For each partition, a score DDSM dataset by the expert radiologists are enlisted in Table 2. 713
    686 of classifier is recorded separately of all partitions and average The single view CAD system detected (164 + 246) 410 lesions 714
    687 score is calculated. This approach of classification is similar to (198 on CC and 212 on MLO view) with image-based detection 715
    5. Experimental results 5.2. Image-based diagnosis 718
    690 5.1. Image-based detection Image-based diagnosis includes classifying the 241 TP lesions 719
    as either benign mass or malignant tumour. For the single view 720
    691 As a first part of our experiment, the individual images of CADx system, a total of 72 textural features and 7 geometric 721
    692 ipsilateral views are pre-processed and then segmented to features are extracted from all the lesions detected as 722
    693 identify the suspicious lesions. Some of these identified suspicious. The performance of the classifier is tested using 723
    694 lesions which are actually not true are known as false combined features only. Feature vectors of all 241 TP 724
    695 positives (FP). The total number of FPs from all the images suspicious lesions are divided into 4 sets of 48 lesions for 725
    696 divided by number of images give rise to the term false training and remaining set of 49 for testing in a five-fold cross 726
    697 positives per image (FPs/I). One of the goals in this article is validation using SVM classifier. Five folds allow more variety of 727
    698 to reduce the FPs/I to the extent possible. A set of empirical