Archives

  • 2019-07
  • 2019-08
  • 2019-09
  • 2019-10
  • 2020-03
  • 2020-07
  • 2020-08
  • 2021-03
  • br Spectral Coherence br MI Mutual

    2019-09-16

    -
    % Spectral Coherence -
    MI Mutual Information -
    System Identification
    Pe Peclet Number -
    v Velocity
    mm
    3-D Convective-Dispersion Modelling
    vCD Convective Velocity mm s 1
    mm
    Heuristic Parameters
    WIT Wash-In Time s
    WIR Wash-In Rate a.u. s 1
    PI Peak Intensity a.u.
    AT Appearance Time s
    Fractal Dimension Analysis
    FD Fractal Dimension -
    PCa = prostate cancer; sPCa = significant prostate cancer; TIC = time-intensity curve.
    The highest performances for PCa and sPCa are indicated in boldface.
    Table 2. Leave-one-patient-out cross-validated ROC-AUC performance of vCD, WIT and a multi-parametric approach for both benign (disease) versus PCa and benign (disease) versus sPCa
    Parameter PCa sPCa
    GMM
    GMM = Gaussian mixture model; PCa = prostate cancer; ROC-AUC = area under the receiver operating characteristic curve; sPCa = significant prostate cancer; SVM = support vector machine; vCD = convective velocity; WIT = wash-in time.
    The highest performances for PCa and sPCa are indicated in boldface.
    growth factor compared with BPH alone (Huang et al. 2018), for example. This might explain why especially regions whose biopsy cores contain prostatitis seem to mimic those with PCa. For this study, it should be under-lined that the 17 cores carrying prostatitis were taken from only four patients in total. It is also worth mentioning that prostatic α-CEHC is considered a distinctive feature for risk of PCa development (Sandhu 2008).
    Another trend discernible for almost all parameters was a persistent difference in variance among regions with iPCa 
    and sPCa. Whereas sPCa is strongly characterized by ele-vated levels of r, Pe and vCD, for example, and reduced WIT and m, a number of iPCa regions exhibit values comparable to those containing BPH or no disease at all. As iPCa is gen-erally not associated with life-threatening risks, showing a very low probability of metastasis (Kryvenko and Epstein 2015; Ross et al. 2012), this might not pose a problem. The management of Gleason 3 + 3 disease in the diagnostic pro-cess and treatment strategy, however, is still a matter of debate among urologists (Mottet et al. 2017). In a more extended set, a grade-weighted or three-class classification approach (i.e., B, iPCa and sPCa) could be considered.
    In this paper, we also evaluated the feasibility of an SVM-based and a GMM-based machine learning approach to classify biopsy regions. The multi-parametric perfor-mance based on GMM outperforms single parameters. Likely, this is because the parameters that are consecutively selected are different in underlying physical meaning; whereas k, r and MI are related to UCA dispersion, m, a, vCD and WIT reflect perfusion. The single-parametric eval-uation as well as the feature selection order (see Fig. 5) showed that flow-related parameters (i.e., vCD and WIT) contribute most to the classification. Interestingly, a domi-nantly featured in the best performing multi-parametric sets even though its individual performance was poor. This might indicate that the number of microbubbles passing a certain voxel (i.e., a) only became important once the TIC shape (i.e., m or WIT) and dispersive fingerprint (i.e., r) were known. Another interesting observation was that, in general, parameters from different analyses were com-bined. These parameter differed not only in physical quan-tities that were estimated or the underlying models that were exploited but also in the scale these processes were assumed to be present. Hence, these analyses could be regarded to carry complementary information useful for the detection of PCa.