Archives

  • 2022-09
  • 2022-08
  • 2022-07
  • 2022-06
  • 2022-05
  • 2022-04
  • 2021-03
  • 2020-08
  • 2020-07
  • 2020-03
  • 2019-11
  • 2019-10
  • 2019-09
  • 2019-08
  • 2019-07
  • br The second aspect is

    2022-04-28


    The second aspect is addressed to indicate some correlations between the best features and the histological structures. For in-stance, the results presented herein allow for the reaching of in-terpretations concerning the attributes and possible relationships
    Table 5
    Breakdown of the selected features using the proposed method.
    FDp
    FDm
    Lac
    Har
    Perc
    Total
    Image H & E Sub image Image H & E Sub image Image H & E Sub image Image H & E Sub image Image H & E Sub image
    linked to structures present on the images. Through observation of the values indicated on Table 6, one can verify that the MP of percolation is less for those images from the benign group than for the malignant group. The AR was also lower in the benign group. The increase in the values of these attributes, as observed between the benign and malignant groups is associated with the presence of cancer. This disease causes an uncontrolled increase of FLAG tag Peptide and degrades the cell structures, such as in lumen regions, intestinal crypts and calceiform cells. The uncontrolled increase of cells can generate more information in the quantification process, which explains the increase in previously indicated attribute val- 
    ues. In Fig. 10 are shown examples of preserved structures from the benign group. The regions degraded by the presence of cancer, malignant group, can be seen in Fig. 11.
    The most relevant set of features was given as input for the classifiers chosen for this study. The results are presented on Table 7. All the classification methods provide differentiation rates above 0.820. Noteworthy here is that the features were su cient to produce relevant performance rates in different classifiers, with av-erage rates of AUC higher than 0.94, such as RaF, AUC = 0.945, K∗ , AUC = 0.947, and PL, AUC = 0.994. This fact indicates the relevance of the discriminative power of the association of features presented
    Table 6
    Features selected at every fold from the k-folds cross validation approach with averages for the benign and malignant groups. The metrics for lacunarity and percolation were area under curve (ARC), skewness (SKW), area ratio (AR), maximum point (MP) and the scale of the maximum point (SMP).
    Feature Image Benign Malignant
    Fig. 9. Average rates for AUC calculated from the folds and applying the methods DT, SVM, NB, RaF, K∗ and PL. The values were obtained with subsets of 1 to 100 features, limits applied to each fold.
    Fig. 10. Visualisation of a benign group with regions marked in blue to indicate the preservation of the organisation and the distribution of crypts (a) and (b), as well as goblet cells (c). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
    Fig. 11. Image of the malignant group with regions marked in blue to indicate the degradation of the organisation and the distribution of crypts (a) and (b), as well as goblet cells (c). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
    Table 7
    Performances for the classification algo-
    rithms with the associations of features
    obtained in our model.
    Classifier AUC
    in our proposal. From such highlights, one can conclude that the best association was reached with the fractal attributes of lacu- 
    narity and percolation multiscale and multidimensional, calculated from curvelet sub-images by applying the PL classifier (Do Nasci-mento et al., 2013).
    5.1. Analysis of the impact of noise on different features
    The best feature and classifier association for distinguishing col-orectal cancer was tested with the addition of different levels of noise on the feature sets. This procedure allowed for the analysis of the impact that this process had on the robustness of the fea-tures and the proposed association. The noise rates were 20% and 40%, as discussed by Zhu and Wu (2004) and Xiao et al. (2010). For each dataset, consideration was given to the AUC average with k-fold cross-validation. In Fig. 12 the average rates are presented for AUC arising from the best association: fractal attributes from
    Table 8
    Performances from the proposed method and correlated studies considering the values of AUC.
    Number of
    Reference features Feature extraction Pre-processing step Classifier AUC
    Masood and 59 Circular LBP features No SVM 0.900 Rajpoot (2009)
    Regression
    porcentage cluster area features
    Boost
    Fourier descriptors and