4377

Detection of  Clinically Significant Prostate Cancer: Incremental Value of Deep Learning to PI-RADS V2
Liang Wang1

1Department of Radiology, Tongji Hospital,Tongji Medical College,Huazhong University of Science & Technology, Wuhan, China

Synopsis

Deep learning has great potential in medical imaging. 168 patients underwent 3T mpMRI of prostate before mpMR-targeted biopsies plus systematic sampling. Two radiologists from two separate institutions, by using the Prostate Imaging Reporting and Data System (PI-RADS) V2 and a multimodal convolutional neural networks (CNN)-based deep learning, independently assessed prostate MRI examinations. Histopathologic findings were used as the reference standard. In detecting csPCa, both reviewers had significantly higher AUCs using CNN-based deep learning. Reviewer 2 benefited much more from CNN-based deep learning than did reviewer 1. Combined PI-RADS with CNN-based deep learning contribute significant incremental value in the detection of csPCa.

INTRODUCTION

Deep learning techniques great potential in medical imaging. The potential of features previously unknown or ignored by humans has been recently explored by automatically learned from a large set of training data based on the state-of-the-art deep convolutional neural networks (CNNs). The purpose of our study to determine whether use of the state-of-the-art CNN-based deep learning improves csPCa with mpMRI when pathologic findings are the reference standard.

METHODS

168 patients underwent 3T mpMRI of the prostate before mpMR-targeted biopsies with cognitive transrectal US-MR fusion plus concurrent 12-core systematic sampling. Two radiologists from two separate institutions, by using the Prostate Imaging Reporting and Data System (PI-RADS) V2 lexicon and a multimodal CNN-based deep learning, independently assessed prostate MR imaging examinations performed at a single center to score the risks of csPCa, blinded to the all clinical and pathologic findings, respectively. A prediction value of combined the predictions of PI-RADS with the predictions of CNN-based deep learning for each patient was obtained by using binary logistic regression model. Histopathologic findings were used as the reference standard. Model performance was compared by receiver operating characteristic (ROC) curve analysis and calculation of areas under the ROC curves (AUCs) and weighted kappa statistics. P < 0.05 was considered significant for all comparisons.

RESULTS

At histologic examination, 64 (38 %) of the patients were found to have csPCa. In detecting csPCa, both reviewers had a significantly higher AUC using CNN-based deep learning (p < 0.001 for both). Reviewer 2 benefited much more from CNN-based deep learning than did reviewer 1. The weighted kappa value was 0.638 for PI-RADS alone and 0.833 for PI-RADS with CNN-based deep learning, indicating fair to good interobserver agreement. Sensitivity and specificity with PI-RADS and with CNN-based deep learning respectively, were 93.8% and 91.3% and 96.9% and 90.4% for reviewer 1 and 71.9% and 84.6% and 82.8% and 92.3% for reviewer 2. The incremental effect of CNN-based deep learning to PI-RADS V2 stratified by the likelihood of csPCa was 92% for Reviewer 1 and 88% for Reviewer 2.

DISCUSSION

mpMRI is generally considered the most accurate imaging method for csPCa [1] . Futterer et al. [2] conducted conducted a systematic review of the literature of 12 studies from January 1, 2000 to September 30, 2014, the findings showed that mpMRI had a sensitivity of 58–96% and specificity of 23–87% in detection of csPCa.

The role of mpMRI in prostate cancer management has been changing with the development of techniques such as deep learning, which is part of a broader family of artificial intelligence and machine learning [3,4]. Several studies have been successfully conducted to explore the ability of CNN-based deep learning and achieved remarkable results for various medical applications, such as thoraco-abdominal lymph node detection in computed tomography (CT) scans, interstitial lung disease classification in CT scans, (39)polyp detection in colonoscopy videos,pulmonary nodule detection in CT images(41).The use of a deep learning simplifies workflow, enhances productivity. Improvements in patient care include shorter reading time, shorter hospital stays, decreased report-waiting times, faster diagnoses, decreased inter-/intra-observer variations.

Our study showed that in the detection of csPCa, radiologists performed substantially better using mbined PI-RADS V2 with CNN-based deep learning than with PI-RADS V2 alone. CNN-based deep learning is particularly helpful in patients with PIRADS 3 (the presence of csPCa is equivocal), CNN-based deep learning changed the csPCa judgment in 16 of 17 patients (94%) for Reviewer 1 and in 76 of 77 patients (99%) for Reviewer 2; in 13 of these 16 patients (81%) for Reviewer 1 and in 67of 76 patients (88%) for Reviewer 2, the changes were appropriate.

CONCLUSION

CNN-based deep learning allows radiologists to more accurately interpret MR images of the prostate, significantly improving csPCa with MRI. Some reviewers benefit more than others from use of this tool

Acknowledgements

N/A

References

[1] Weinreb JC, Barentsz JO, Choyke PL, Cornud F, Haider MA, Macura KJ, Margolis D, Schnall MD, Shtern F, Tempany CM et al: PI-RADS Prostate Imaging - Reporting and Data System: 2015, Version 2. European urology 2016, 69(1):16-40.

[2] Venderink W, van Luijtelaar A, Bomers JG, van der Leest M, Hulsbergen-van de Kaa C, Barentsz JO, Sedelaar JP, Futterer JJ: Results of Targeted Biopsy in Men with Magnetic Resonance Imaging Lesions Classified Equivocal, Likely or Highly Likely to Be Clinically Significant Prostate Cancer. European urology 2017

[3] Fehr D, Veeraraghavan H, Wibmer A, Gondo T, Matsumoto K, Vargas HA, Sala E, Hricak H, Deasy JO: Automatic classification of prostate cancer Gleason scores from multiparametric magnetic resonance images. Proceedings of the National Academy of Sciences of the United States of America 2015, 112(46):E6265-6273.

[4] Peng Y, Jiang Y, Yang C, Brown JB, Antic T, Sethi I, Schmid-Tannwald C, Giger ML, Eggener SE, Oto A: 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, 267(3):787-796.

Figures

Figure 1 illustrates the end-to-end proposed framework for prediction of identification the presence of csPCa in a patient

Figure 2: Graph shows results of receiver operating characteristic analysis for detection of csPCa with PI-RADS (by Reviewer 1) alone and PI-RADS with CNN-based deep learning

Figure 3: Graph shows results of receiver operating characteristic analysis for detection of csPCa with PI-RADS (by Reviewer 2) alone and PI-RADS with CNN-based deep learning

Proc. Intl. Soc. Mag. Reson. Med. 26 (2018)
4377