Jing Zhang^{1}, Yu-dong Zhang^{2}, Yang Song^{1}, Xu Yan^{3}, and Guang Yang^{1}

Timely diagnosis and treatment could effectively reduce patient risk for clinical significant prostate cancer (PCa). In this abstract, we extracted 327 quantitative features from prostate mp-MRI images, then we used a homemade open-source tool named Feature Explorer to study combinations of radiomics algorithms and hyper-parameters in order to find the best model for classification of PCa into non-clinical–significant and clinical significant. We obtained a candidate model with AUC of 0.823, accuracy of 0.827. Four features selected for classification are easily understandable in the sense of image characteristics. Feature Explorer was demonstrated to be an efficient tool for radiomics studies.

**Dataset:** PROSTATEx
(https://doi.org/10.7937/K9TCIA.2017.MURS5CL) dataset
was used in this study. It included 185 cases with T2W(TSE，0.5×0.5×3.6mm^{3})， DWI(SSEP，2×2×3.6mm^{3}，b is
800s/mm^{2}), and ADC map sequences from Siemens 3T MR scanners. Total 251
lesions (CS/NCS=68/183) were used in this study. DWI and ADC map were aligned to
T2W images. A radiologist drew the region of interest (ROI) manually. We split
the dataset into independent training (CS/NCS = 48/128) and testing dataset (CS/NCS
= 20/55).

**Radiomics Feature Extraction:** We extracted
109 features from each ROI in each sequence with pyradiomics (http://pyradiomics.readthedocs.io/en/latest/index.html). Classes
of the features used included Shape (19), First Order (16), Gray Level
Co-occurrence Matrix (GLCM, 23), Gray Level Size Zone Matrix (GLSZM, 16), Gray
Level Run Length Matrix (GLRLM, 16), Neighboring Gray Tone Difference Matrix
(NGTDM, 5), Gray Level Dependence Matrix (GLDM, 14).

**Feature
Explore Pipeline: **Since there are numerous number of combinations of
algorithms and hyper-parameters to try out to find the best model for
classification, we used a homemade open-source tool named Feature Explorer
(FAE, https://github.com/salan668/FAE) to automate
the process. We normalized each features, and used upsampling for data balance.
Then we tried out all the combinations of three feature selection methods (ANOVA,
Relief, and Recursive feature elimination) and four classifiers (SVM, LDA,
Logistic Regression, and Logistic Regression with Lasso). Number of selected
features was also iterated from 1 to 20. The best model was found by comparing
the results of leave-one-out cross validation on the training dataset. Finally,
we used receiver operating characteristic curve (ROC), area under ROC (AUC),
paired t-test on the testing dataset to quantitatively evaluate the performance
of the best model.

We found
that the combination of ANOVA and LDA with 4 features selected yielded the best
results, with AUC of 0.823, accuracy of 0.827, sensitivity of 0.800, specificity
of 0.836, positive predictive value of 0.640, negative predictive value of 0.920.
We showed the ROC curve of the model on training/validation/testing dataset in
Figure 2 (a). The plot of the AUC on validation dataset against the number of
features was shown in Figure 2 (b). The candidate number of features was determined
with one-standard-error rule. The selected features were: (1) 10th percentile of ADC map (10Per-ADC),
(2) the interquartile range of intensity analysis of DWI (IR-DWI), (3) auto-correlation
of GLCM of DWI (AC-GLCM-DWI), and (4) the gray level variance of GLSZM of DWI
(GLV-GLSZM-DWI). The contributions of these four features in the final model
were shown in
Figure
2. We also showed
the distribution of these features of both CS and NCS PCa in Figure 3. The
p-value of these features was smaller than 0.001 to distinguish the CS and NCS
PCa. The histogram within ROI of CS and NCS PCa cases were shown in Figure 4.
The features related to the histogram could be used to separate the CS and NCS
PCa.

1.Canadian Cancer Society, Prostate Cancer Statistics, 2015.

2.American Cancer Society, Cancer Facts & Figures 2015.

3.A. Jemal, F. Bray, M. M. Center, J. Ferlay, E. Ward, and D. Forman, CA: a cancer journal for clinicians: Global cancer statistics, 2011; vol. 61, no. 2, pp. 69–90.

4.A. Stangelberger, M. Waldert, and B. Djavan, Prostate cancer in elderly men, Rev. Urol, 2008; vol. 10, no. 2, pp. 111–119.

5.Y. Peng et al. Quantitative analysis of multiparametric prostate MR images: Differentiation between prostate cancer and normal tissue and correlation with Gleason score—A computer-aided diagnosis development study, Radiology, 2013; vol. 267, no. 3, pp. 787–796.

6.P. Lambin, E. Rios-Velazquez, R. Leijenaar, S. Carvalho, et al. Radiomics: extracting more information from medical images using advanced feature analysis, European Journal of Cancer, 2012; vol. 48, no. 4, pp. 441–446.

Figure 1 : Radiomics analysis flowchart. We normalized gray
level in all images before feature extraction. Features were extracted with pyradiomics.
All those extracted features will be filtered with cosine-similarity before
feature selection algorithms were used to select relevant features. Survived
features were then used to train the classifier. All the processing after
feature extraction were performed in FeatureExplorer.

Figure 2: The optimal model found by FAE, in which ANOVA is
used for feature selection, and LDA as classifier. Four features were selected
by the model. (a) ROC curve of the optimal model with Train AUC of 0.838, Validation
AUC of 0.814, Test AUC of 0.824. (b) The plot of the AUC value against the
number of features. We used one-standard-error rule to find the candidate
number of features. (c) Contributions of the selected features, namely 10Per-ADC
(weight value in LDA is -7.98), IR-DWI (weight value in LDA is 1.89), AC-GLCM-DWI
(weight value in LDA is 0.95), GLV-GLSZM-DWI (weight value in LDA is 14.96).

Figure 3.The histograms of the 4
most important features: The blue and green bars denote NCS and CS PCa,
respectively. It can be seen that 10Per-ADC tends to have lower values in CS
while IR-DWI, AC-GLCM-DWI, and GLV-GLSZM-DWI tend to have higher values in CS.
The p-values of all 4 features to distinguish CS/NCS are less than 0.001.

Figure 4. CS/NCS images and corresponding histograms of ROI. As
seen in ADC maps in (a) (b), cancer region features higher signal in NCS and histograms
of NCS and CS show opposite skewness. As seen from DWI images in (c) (d),
cancer region in NCS has lower intensities and narrower distribution in histogram.