Dharmesh Singh1, Virendra Kumar2, Chandan J Das3, Anup Singh1, and Amit Mehndiratta1
1Centre for Biomedical Engineering (CBME), Indian Institute of Technology (IIT) Delhi, New Delhi, India, 2Department of NMR, All India Institute of Medical Sciences (AIIMS) Delhi, New Delhi, India, 3Department of Radiology, All India Institute of Medical Sciences (AIIMS) Delhi, New Delhi, India
Synopsis
Automatic grading of prostate cancer
(PCa) can play a major role in its early diagnosis, which has a significant impact
on patient survival rates. The objective of this study was to develop and
validate a framework for classification of PCa grades using texture features of
T2-weighted MR images. Evaluation of classification result shows accuracy of
85.10 ± 2.43% using random forest feature selection and Gaussian support-vector
machine classifier.
Introduction
Early and accurate detection of
clinically significant prostate cancer (PCa), which is defined as grade-2 or
higher1, is essential to improve the patient’s outcomes. MRI has
become integral diagnostic procedure for the detection and risk stratification
of PCa2. MRI includes image sequences such as T2-weighted imaging
(T2WI), diffusion-weighted imaging (DWI) with derived apparent-diffusion
coefficient (ADC) map and dynamic contrast-enhanced (DCE) imaging2.
According to Prostate Imaging-Reporting and Data System version-2 (PI-RADS v2),
DWI and ADC are useful for the assessment of peripheral-zone (PZ) lesions and
T2WI is useful for the assessment of transition-zone (TZ) lesions3. PI-RADS
v2 assesses the likelihood of PCa on a 5-point scale for each lesion, with
grade-1 and grade-2 are considered as low-grade (LG), grade-3 as intermediate-grade
(IG), whereas grade-4 and grade-5 as high-grade (HG)3. Texture-analysis
based machine-learning (ML) models could adjunct the clinical workflow and
increase efficiency to increase accurate tumor evaluation and characterization4.
Different classification models using MRI based texture features have been
presented in the literature5-7. The aim of this study was to
investigate the role of T2W MRI based texture features in characterization of PI-RADS
v2 grades of PCa. Methods
MRI
data acquisition:
MRI dataset of 59 patients (age:65.0 ±
8.5 years) with clinically proven PCa was used in this retrospective study. All
prostate MRI examinations were acquired at 1.5T scanner (Achieva, Philips Health
Systems,The Netherlands). MRI acquisition sequences included axial turbo-spin
echo (TSE) T2WI (TR/TE=3330/90 ms, field-of-view (FOV)=250×250 mm2, acquisition-matrix=320×320,
voxel-size=0.49×0.49×3 mm3, slice-thickness=3 mm, number-of-slices=36
and DWI sequence was performed using echo-planar imaging (TR/TE=6831/81 ms, FOV=292
×292 mm2, acquisition-matrix=112×112, voxel-size=2.6×2.6×3 mm3,
slice-thickness=3 mm, number-of-slices=36, with five b-values of 0,500,1000,1500
and 2000 s/mm2). ADC maps were generated using all five b-values
with the least square-optimization to the mono-exponential model using the
vendor-provided algorithm at the clinical workstation8.
Methodology:
In the
pre-processing step, co-registration and manual prostate gland and zonal segmentation
were done. The 3D affine-transformation method with mutual-information
similarity index was used for co-registration of T2WI and DWI. A total of 435 lesion ROIs
(size:50-200 voxels) marked according to
PI-RADS v2 guidelines3 were manually delineated from segmented PZ of T2WI,
DWI and ADC each with
the help of a radiologist (with>10 years of experience in prostate-imaging).
In order to differentiate the different aggressiveness of PCa, a set of
features was calculated. The six texture feature models were consisted of a
total of 49 features for each imaging modality, as presented in table-1. Out of 49 features, some
may turn out to be less informative. Random-forest (RF) and
correlation-based feature selection (CFS) methods were used for selecting the best
features. For PCa classification, support vector machine (SVM; kernel=linear and Gaussian) and K-nearest
neighbour (KNN; K=10) were used. The first step of the framework was to classify the lesions into
three-classes (LG vs. IG vs. HG) and second step to sub-classification of HG
into two-classes (PI-RADS grade-4 vs. grade-5).
Data were processed using in-house developed routine in MATLAB® (MathWorks,v2018,Natick). The performance of the proposed methodology was evaluated using
10-fold cross-validation (CV) and measured using sensitivity, specificity,
accuracy and area under the receiver-operating characteristics (ROC) curve (AUC).Results
Using T2W MRI,
the optimal texture features from CFS method and Gaussian SVM classifier
achieved best performance with sensitivity,68.86±1.18%, specificity,89.07±0.14%,
accuracy,73.55±0.35% and AUC,0.90 for three-class: LG vs. IG vs. HG classification
and feature-set from CFS and linear SVM achieved best performance with
sensitivity,72.91±0.67%, specificity,91.64±0.91%, accuracy,75.40±0.57% and
AUC,0.83 for two-class: grade-4 vs. grade-5 classification. The combined T2WI, DWI
and ADC texture features from RF feature-selection and Gaussian SVM classifier
achieved the accuracy, 91.85±0.49% for three-class and 90.55±0.64% for
two-class classification. The cumulative performance of the two-step was
evaluated for classification of PI-RADS v2 grades (2-5) and achieved accuracy,61.60+1.28%
using T2WI and 85.10±2.43% using a combination of T2WI, DWI and ADC. The
performance of the classification using different classifiers is presented in
table-2 and table-3. Figure-1 represents the ROC plots for LG vs. IG vs. HG and
grade-4 vs. grade-5 classifications using T2WI.Discussion
The
study evaluated the diagnostic performance of T2WI based texture features in characterization
of PI-RADS v2 grades of PCa. T2WI showed accuracy, 73.55% for LG vs.
IG vs. HG and 75.40% for PI-RADS grade-4 vs. grade-5 classification. In the previous
studies, automated detection and classification systems have been presented
based on T2WI9 and mpMRI5-7 with pooled accuracy,86% (73-89%).
The proposed framework was found to have accuracy of 61.60% for classifying lesions
into PI-RADS v2 grades (2 to 5) using T2WI. In this study, T2WI showed low accuracy in
classification of PCa same as has been suggested in the literature10,11.
DWI and ADC have been demonstrated to add a lot of value in characterization of
prostate tumor12 same has been observed in our study where the
classification performance increased the accuracy by more than 23% (from 61.60%
to 85.10%) by combining T2WI and diffusion MRI.Conclusion
The
proposed study suggests that the use of texture features extracted from T2WI, DWI
and ADC improve the accuracy of PCa characterization by almost 23% compared
to T2WI alone. The proposed framework obtained classification accuracy of 85.10%
for PCa grades based on PI-RADS v2, which can improve the diagnostic performance
for prostate cancer treatment.Acknowledgements
This
work is supported by IIT Delhi, India and AIIMS New-Delhi, India. DS was
supported with the research fellowship fund from the Ministry of Human Resource
Development, Government of India.References
- Carter HB, Partin AW, Walsh PC, et al. Gleason score 6 adenocarcinoma:
Should it be labeled as cancer. J.
Clin. Oncol. 2012;30:4294-4296.
- Rooij MD, Hamoen EHJ, Fütterer JJ, et al. Accuracy of multiparametric
MRI for prostate cancer detection: A meta-analysis. Am. J. Roentgenol. 2014;202:343-351.
- Weinreb JC, Barentsz JO, Choykeet PL, et al. PI-RADS Prostate Imaging -
Reporting and Data System: 2015, Version 2. Eur Urol. 2016;69(1):16-40.
- Nelson CR, Ekberg J, Fridell K. Prostate Cancer Detection in Screening
Using Magnetic Resonance Imaging and Artificial Intelligence. The Open
Artificial Intelligence Journal. 2020;6:1-11.
- Shah V, Turkbey B, Mani H, Pang Y, et al.
Decision support system for localizing prostate cancer based on multiparametric
magnetic resonance imaging. Med Phys. 2012; 39: 4093-4103.
- Kwak JT, Xu S, Wood BJ, et al. Automated
prostate cancer detection using T2-weighted and high-b-value diffusion-weighted
magnetic resonance imaging. Med Phys. 2015;42:2368-2378.
- Khalvati F, Wong A, Haider MA. Automated
Prostate Cancer Detection via Comprehensive Multi-parametric Magnetic Resonance
Imaging Texture Feature Models. BMC Med. Imaging. 2015;15(1):27.
- Peng Y, Jiang Y, Antic T, et al. Apparent diffusion coefficient for
prostate cancer imaging: impact of B values. AJR Am J Roentgenol. 2014;202(3):W247-W253.
- Zhao K, Wang CY, Hu J, et al. Prostate cancer identification:
quantitative analysis of T2-weighted MR images based on a back propagation
artificial neural network model. Sci China Life Sci, 2015;58:666-673.
- Toivonen J, Montoya Perez I, Movahedi P, et al. Radiomics and Machine
Learning of Multisequence Multiparametric Prostate MRI: Towards Improved
Non-Invasive Prostate Cancer Characterization. PLoS ONE. 2019;14:e0217702.
- Rudolph MM, Baur ADJ,
Haas M et al. Validation of the PI-RADS language: predictive values of PI-RADS
lexicon descriptors for detection of prostate cancer. Eur Radiol.
2020;30:4262-4271.
- Wibmer AG, Sala E, Hricak H, et al. The expanding landscape of
diffusion-weighted MRI in prostate cancer. Abdom Radiol. 2016;41:854-861.