Introduction

Multiparametric-MRI (mpMRI) based prostate screening method is able to improve the early detection of high-risk prostate cancer (PCa).¹ Recently, Prostate Imaging-Reporting and Data System version-2.1 (PI-RADS v2.1) was designed to improve global standardization process and reduce heterogeneity in prostate mpMRI acquisition, interpretation, and reporting.² PI-RADS assessment system is a qualitative Likert-based [1 to 5] scale with higher values indicating higher suspicion of PCa.² PI-RADS includes a lesion-size based decision-criterion (cut-off:1.5cm for scores-4 and 5) and also provides a minimal requirement for the measurement of lesion-volume (>0.5cc for high-risk PCa).² One study by Martorana et al. reported the PI-RADS scores increase with the increase in lesion-volume.³
However, PI-RADS scoring is challenging due to its inherent technical difficulties to visualize a small lesion on MRI, which leads to subjectivity and the whole PI-RADS process is time-consuming as reporting time has become an important performance indicator in healthcare.⁴ Previous studies have shown poor inter-reader variability in the assessment of PI-RADS scores.^5,6 Machine learning-based automated or semi-automated PI-RADS scoring has the potential to assist radiologists in screening process, reducing inter-reader variability and evaluation-time.⁷ The objective of this study was to develop a semi-automated framework to simplify the reporting process of prostate MRI and validate the performance of framework using machine-learning methods.

Methods

MRI data-acquisition
MRI dataset of 59 men (mean-age:65 ± 8.5 years) with clinically proven PCa (PI-RADS v2.1 score-2=16, score-3=10, score-4=18 and score-5=15) was used in this retrospective study. All prostate MRI examinations were acquired at 1.5T scanner (Achieva,Philips Health Systems). T2-weighted images were acquired using a turbo-spin-echo sequence with 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 and number-of-slices=36. Diffusion-weighted images were acquired 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-slice=36, with five b-values of 0,500,1000,1500 and 2000 s/mm²). Apparent-diffusion coefficients were calculated using the vendor-provided software at the clinical-workstation.
Methodology
The pre-processing steps involved prostate-gland segmentation, 3D image-registration, prostate-zonal segmentation, lesion region-of-interest (ROI) marking and lesion measurement. The Chan-Vese active-contour model along with morphological opening operation was used for prostate-gland segmentation and a probabilistic atlas with partial-volume (PV) correction algorithm was used for prostate-zonal segmentation.⁸ An affine-transformation method with a mutual-information similarity index was applied for 3D-registration of the prostate gland ROIs of T2WI and DWI.
Lesion ROIs were marked on the peripheral-zone of DWI (b=2000 s/mm²), ADC and T2W images as per PI-RADS v2.1 guidelines² with the help of an expert radiologist. The ellipse-fitting approach was used for the measurement of lesion-maximum diameter and volume. PI-RADS v2.1 scores were assessed based on lesion measurements using ellipse-fitting. Linear discriminant-analysis (LDA), linear support-vector machine (SVM) and Gaussian SVM were used to evaluate the diagnostic accuracy of the proposed framework by classification of PI-RADS scores; i)score-2 vs. score-3 vs. score-4 vs. score-5 and ii)low-score (score 2,3) vs. high-score (score 4,5) using stratified 5-fold cross-validation (CV).
Data were processed using in-house developed toolbox in MATLAB (v.2018;MathWorks,Natick,USA). Sensitivity, specificity, accuracy and area under the receiver-operating characteristic curve (AU-ROC) were measured to evaluate the performance of the proposed framework. The workflow of our proposed framework is shown in figure-1.

Results

Lesion-maximum diameter and lesion-volume measured by the ellipse-fitting approach were 0.47±0.06 cm and 0.13±0.03 cc for score-2; 0.67±0.11 cm and 0.25±0.12 cc for score-3; 0.96±0.18 cm and 0.56±0.28 cc for score-4 and 1.45±0.15 cm and 0.99±0.25 cc for score-5, respectively. The proposed framework-based PI-RADS v2 assessment showed 50 out of 59 subjects correctly matched (~85%) with the radiologist assessment. The proportion of correct classification rate in all PI-RADS scores is shown in figure-2. Semi-automated PI-RADS assessment showed strong positive-correlation (r=0.94,p<0.05) with radiologist-assessment.
Table-1 presents the performance of both approaches, i)four-class classification (score-2 vs. score-3 vs. score-4 vs. score-5) and ii)two-class (low-score vs. high-score) using three different classifiers. LDA classifier achieved the highest performance, sensitivity,85.50±1.95%; specificity,75.00±1.10%; accuracy,88±0.98%; AUC,0.94 in four-class classification and linear SVM classifier achieved the highest performance, sensitivity,91.45±3.65%; specificity,95.85±1.25%; accuracy,93.20±2.10%; AUC,0.99 in two-class classification using 5-fold CV. Figure-3 shows the ROC graphs for four-class and two-class classifications using different classifiers.

Discussion

An automatic or semi-automatic PI-RADS scoring could assist the radiologist to speed up reporting and reduce errors in misclassifying lesions. PI-RADS scoring has inherent subjectivity due to its reliance on interpretable descriptions and lack of quantitative metrics.² Previously, the performance of semi-automated PI-RADS assessment has been evaluated using a convolution-neural network with the mean AUC of 0.65 and both studies were based on two-class (low vs. high score) classification.^9,10 The proposed framework here outperformed the existing methods with the correct classification rate of 85% and AUC,0.94 for four-class classification (score-2 vs. score-3 vs. score-4 vs. score-5) and AUC,0.99 for low-score (score 2&3) vs. high-score (score 4&5) classification. The proposed framework for semi-automatic PI-RADS v2.1 scoring is relatively objective and could be helpful for non-expert radiologists in terms of reporting accuracy. Future work will include automated ROI-lesion delineation and a large number of patients from multiple clinical centres to evaluate the accuracy of proposed framework.

Conclusion

The developed semi-automatic framework for PI-RADS v2.1 scoring achieved high classification accuracy (88%) using machine-learning approach. This framework could improve the consistency of scoring in a screening setting of prostate cancer.

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.