• 2019-07
  • 2019-08
  • 2019-09
  • 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • Because of the limited number of


    Because of the limited number of patients in our data set, the learning was restricted to the region-mean values of the parameters. In radiologic practice, how-ever, the spatial appearance of lesions, such as tissue het-erogeneity or abnormal spiculation, is also often assessed. Radiomic features allow quantitative extrac-tion of temporal and spatial information from (paramet-ric) images and look beyond the plain voxel-specific parameter or gray-scale values (Aerts et al. 2014; Lam-bin et al. 2012; Larue et al. 2017). Therefore, the exten-sive set of features described in this Spectinomycin work (such as the maximum value or entropy) might become very useful when applying machine learning in a more extended data set. The same holds for information from other vas-cular US imaging techniques (e.g., molecular imaging [Turco et al. 2017], ultrafast Doppler [Provost et al. 2015], super-resolution US [Couture et al. 2018] or acoustic angiography [Gessner et al. 2013]), other
    Fig. 5. (a) Parameter selection frequency over all cross-validation folds shown as bar height with the selection order depicted in color; (b) The GMM score distribution over benign and malignant diseases, similar to Figure 4. a = area under the time-intensity curve; AT = appearance time; B = benign; BPH = benign prostatic hyperplasia; DCD = convec-tive dispersion; D = dispersion; FD = fractal dimension; GMM = Gaussian mixture model; iPCa = insignificant prostate cancer; k = dispersion parameter; MI = mutual information; P = prostatitis; Pe = Peclet number; PI = peak intensity; r = spatiotemporal correlation; sPCa = significant prostate cancer; m = mean transit time; v = velocity; vCD = convective velocity; WIR = wash-in rate; WIT = wash-in time; % = spectral coherence.
    imaging modalities or even blood tests and other clinical data that could provide useful complementary features on a patient level.
    Unfortunately, SBx is known to sometimes misrep-resent the actual histopathologic fingerprint because of sampling errors (Eichler et al. 2006; Graif et al. 2007; Ukimura et al. 2013) due either to a mismatch between the template and actual location or to missing the region’s cancer hotspot. This study might therefore wrongly assume regions to be true negative or false posi-tive. As our analysis was based on the entire region surrounding the biopsy location, we expected that in par-ticular some false positive regions would have been upgraded in a more rigorous histologic examination. As these errors in the ground truth could have potentially impacted both the training and testing, their effect might be significant. In the future, a radical-prostatectomy ref-erence study would enable us to assess the presented (multi)parametric maps value for the localization of PCa. However, such a data set would be severely biased toward PCa presence because healthy, BPH-only or prostatitis-only prostates, which coincidentally comprise a large part of the biopsy population, cannot be repre-sented. A prospective targeted-biopsy study with 3-D-CUDI guidance might eventually help determine its PCa-detection ability. In the last two approaches, we envision the use of a 3-D multi-parametric map (similar to those displayed in Fig. 2) rather than biopsy-region-based assessment.
    This study revealed the potential of a multi-parametric combination of 3-D CUDI parameters, either as single parameters or in a multi-parametric approach through 
    machine learning by a Gaussian mixture model. We envis-age additional optimization of 3-D algorithms and expan-sion of the data set to prompt further development of helper T cells technique (e.g., toward deep learning approaches). Eventu-ally, we believe these approaches might allow reliable imaging of PCa for targeted biopsy procedures.