Machine learning classifiers can predict Gleason pattern 4 prostate cancer with greater accuracy than experienced radiologists

Michela Antonelli,Alex Freeman,Clare Allen,King K Cheung,Edward W Johnston,Lucy A M Simmons,Nikolaos Dikaios,Sebastien Ourselin,David Atkinson,Harbir S Sidhu,Hashim U Ahmed,Shonit Punwani,Francesco Giganti,Mrishta B Appayya

doi:10.1007/s00330-019-06244-2

Michela Antonelli, Alex Freeman + Show 12 more

Open Access

https://doi.org/10.1007/s00330-019-06244-2

Copy DOI

Abstract

ObjectiveThe purpose of this study was: To test whether machine learning classifiers for transition zone (TZ) and peripheral zone (PZ) can correctly classify prostate tumors into those with/without a Gleason 4 component, and to compare the performance of the best performing classifiers against the opinion of three board-certified radiologists.MethodsA retrospective analysis of prospectively acquired data was performed at a single center between 2012 and 2015. Inclusion criteria were (i) 3-T mp-MRI compliant with international guidelines, (ii) Likert ≥ 3/5 lesion, (iii) transperineal template ± targeted index lesion biopsy confirming cancer ≥ Gleason 3 + 3. Index lesions from 164 men were analyzed (119 PZ, 45 TZ). Quantitative MRI and clinical features were used and zone-specific machine learning classifiers were constructed. Models were validated using a fivefold cross-validation and a temporally separated patient cohort. Classifier performance was compared against the opinion of three board-certified radiologists.ResultsThe best PZ classifier trained with prostate-specific antigen density, apparent diffusion coefficient (ADC), and maximum enhancement (ME) on DCE-MRI obtained a ROC area under the curve (AUC) of 0.83 following fivefold cross-validation. Diagnostic sensitivity at 50% threshold of specificity was higher for the best PZ model (0.93) when compared with the mean sensitivity of the three radiologists (0.72). The best TZ model used ADC and ME to obtain an AUC of 0.75 following fivefold cross-validation. This achieved higher diagnostic sensitivity at 50% threshold of specificity (0.88) than the mean sensitivity of the three radiologists (0.82).ConclusionsMachine learning classifiers predict Gleason pattern 4 in prostate tumors better than radiologists.Key Points• Predictive models developed from quantitative multiparametric magnetic resonance imaging regarding the characterization of prostate cancer grade should be zone-specific.• Classifiers trained differently for peripheral and transition zone can predict a Gleason 4 component with a higher performance than the subjective opinion of experienced radiologists.• Classifiers would be particularly useful in the context of active surveillance, whereby decisions regarding whether to biopsy are necessitated.

Highlights

Prostate cancer is a heterogeneous disease, with a strong relationship between aggressiveness, as characterized by Gleason grade, and survival [1]
While there is some evidence that the subjective opinion of radiologists interpreting multiparametric MRI can be used to estimate Gleason grade [9], quantitative measurements of signal intensity including normalized T2 signal intensity and apparent diffusion coefficient (ADC) moderately correlate with Gleason grade [10, 11] and have been shown to differ in peripheral zone (PZ) vs. transition zone (TZ) tumors [12]
SMOTE was applied to the training set before using it to train the classifier for all the analyses performed in the PZ cohort

Summary

Introduction

Prostate cancer is a heterogeneous disease, with a strong relationship between aggressiveness, as characterized by Gleason grade, and survival [1]. The concept of Gleason 3 and Gleason 4 tumors representing distinct disease states has emerged [2], due to the different signatures at a genomic level [3] and the distinct survival rates encountered in large long-term follow-up studies [4, 5]. A reliable, quantitative, and non-invasive test to identify patients at risk of aggressive disease (those with a potential Gleason 4 component) would have significant clinical value but does not currently exist. Clinical parameters such as tumor volume [7] and serum prostate-specific density (PSAd) have been shown to correlate with Gleason grade [8]. While there is some evidence that the subjective opinion of radiologists interpreting multiparametric (mp) MRI can be used to estimate Gleason grade [9], quantitative measurements of signal intensity including normalized T2 signal intensity and apparent diffusion coefficient (ADC) moderately correlate with Gleason grade [10, 11] and have been shown to differ in peripheral zone (PZ) vs. transition zone (TZ) tumors [12]

Methods

Results

Conclusion