RPS 307 - Deep learning and radiomics in prostate imaging

Purpose:

To compare the clinical utility of a convolutional neural network (CNN)-based model applied to multiparametric MRI (mpMRI) with a risk model (RM) combining PI-RADSv2 and clinical parameters, for risk assessment of clinically significant prostate cancer (csPCa; ISUP≥grade 2).

Methods and materials:

We retrospectively analysed 499 patients who underwent mpMRI and MRI-targeted biopsy. A 3D-CNN classifier was trained on diffusion-weighted MRIs and ADC maps of 400 patients (training cohort). CNN output probabilities were recalibrated to provide interpretable risk estimates for csPCa. A subset of the training cohort (n=290) was used to build an RM incorporating PI-RADSv2 scores and clinical parameters (prostate volume, PSA density, and age). Model discrimination in the validation cohort (n=99) was compared using the area under the ROC curve (AUC). Clinical usefulness of both models and PI-RADSv2 (cut-off≥4) was assessed using decision curves illustrating the clinical net benefit, which balances the benefits of detecting csPCa against the harms of unnecessary biopsies.

Results:

The CNN’s discrimination (AUC=0.83) was not superior to the RM (AUC=0.83) (p=0.95); however, across the entire range of clinically relevant risk thresholds (5-20%), the CNN had a higher net benefit than PI-RADSv2 and the RM. The CNN showed a clear net benefit at a risk threshold ≥5% for biopsy referral, resulting in a biopsy reduction of 21% (21/99) without missing csPCa (0/37). Applying a risk threshold ≥10% for the CNN and RM resulted in a 25% (25/99) and 7% (7/99) biopsy reduction, missing 3% (1/37) and 0% (0/37) of csPCas, respectively.

Conclusion:

Our CNN model with calibrated risk estimates could reduce biopsies by 21% without missing any csPCa and is potentially superior to a multivariate RM combining PI-RADSv2 and clinical parameters.

Limitations:

A single-centre, retrospective study.

Ethics committee approval

IRB-approved study.

Funding:

OICR; DFG-fellowship[DE 3207/1-1].

Purpose:

To apply a previously developed convolutional neural network (CNN) ensemble to a new set of prostate MRI examinations from consecutive patients for further validation and to examine the influence of different image co-registration approaches on CNN predictions.

Methods and materials:

3 Tesla MRI from 147 consecutive patients with 251 MR lesions (70 sPC positive) were included. DICOM images of T2-weighted, DWI b=1500 s/mm2, and corresponding ADC maps were analysed by a fully automatic analysis pipeline consisting of DWI/ADC to T2W image co-registration utilising three approaches (rigid, affine, and b-spline) prior to evaluation by the CNN ensemble. Maximum CNN tumour map probability above 0.33 (calibrated previously to correspond to PI-RADS>=4 assessment) indicated positive, otherwise regarded as negative. Both individual and combined maximum probability of the three co-registrations was evaluated.

Results:

Sensitivity/specificity/accuracy was 65%/84%/75% for rigid, 71%/73%/72% for b-spline, 73%/72%/71% for affine, and 77%/70%/71% for combined compared to 89%/60%/73% for PI-RADS>=4. Accuracy (p=0.87), sensitivity (p=0.08), and specificity (p=0.36) of CNN predictions on combined co-registration were not significantly different from PI-RADS>=4 assessment.

Conclusion:

The best CNN sensitivity was achieved by combined registration, highlighting the importance of registration on the performance for detection of small lesions. While achieving a lower sensitivity than PI-RADS at higher specificity, the CNN approach showed promising performance. Importantly, lowering the probability threshold of the CNN ensemble could increase the sensitivity more to match the specificity of PI-RADS. However, more data is required to justify adjusting the threshold.

Limitations:

This is a single institution, single scanner CNN model, which requires multi-institutional, multi-scanner validation in the future, however, benefits from standardised conditions in the current setting.

Ethics committee approval

Approved by an ethics committee. Written informed consent waived (S-156/2018).

Funding:

No funding was received for this work.

Purpose:

To determine the diagnostic performance of the texture analysis of prostate MRI for the diagnosis of prostate cancer among PI-RADS 3 lesions.

Methods and materials:

43 patients with at least one PI-RADS 3 lesion on prostate MRI performed between June 2016 and January 2019 were retrospectively included. The reference standard was a pathological analysis of radical prostatectomy specimens or MRI-targeted biopsies. Texture analysis extraction of target lesions and non-lesional prostate glands was performed on axial T2-weighted images and ADC maps using a radiomic software. Lesions were categorised as prostate cancer (Gleason score≥6) and no prostate cancer. Statistical analysis was performed using the generalised linear model (GLM) regression and the discriminant analysis (DA). AUROC with 95% C.I. were calculated to assess the diagnostic performance of standalone features and predictive models for the diagnosis of prostate cancer.

Results:

The analysis of 46 PI-RADS 3 lesions (i.e. 27 [58.7%] no prostate cancers; 19 [41.3%] prostate cancers) revealed 9 and 6 independent texture parameters significantly correlated with the final histopathological results on T2-weighted and ADC maps images (r_pb of 0.274-0.387 and 0.293-0.371), respectively. The resulting GLM and DA predictive models for the diagnosis of prostate cancer yielded an AUROC of 0.775 and 0.779 on T2-weighted images or 0.815 and 0.821 on ADC maps images.

Conclusion:

Texture analysis of PI-RADS 3 lesions on T2-weighted and ADC maps images helps to identify prostate cancer. The good diagnostic performance of the combination of multiple radiomic features for the diagnosis of prostate cancer may help in predicting lesions where aggressive management may be warranted.

Limitations:

Retrospective analysis, only pathologically proven lesions, and small sample size.

Ethics committee approval

IRB-approved study, informed consent was waived.

Funding:

No funding was received for this work.

Purpose:

To explore first-order ADC radiomics metrics other than the ADC mean value (known to be one of the strongest monoparameteris in quantitative assessment) regarding their ability to further increase the predictive value of prostate MRI for detecting clinically significant prostate cancer (csPC).

Methods and materials:

Consecutive patients with a clinical suspicion for csPC examined on a single 3 Tesla MR scanner who underwent MR-guided targeted and systematic biopsy in 2017 and had visible MR lesions were included. A board-certified radiologist re-reviewed MRIs blinded to clinical information and segmented lesions on MRI images. The mean ADC (mADC) and 18 other first-order ADC radiomic parameters were calculated. Logistic regression with variable selection was used to determine the optimal parameters for csPC prediction. Receiver operating characteristics (ROC) and a likelihood-ratio test (LRT) were used for comparison of performance.

Results:

253 patients harboured 392 MR lesions (302 in the peripheral zone), of which 129 lesions were csPC positive. The univariate model ROC area under the curve (AUC) was 0.66 for mADC and 0.68 for entropy. Variable selection retained only mADC and entropy in the final model, while none of the other radiomic parameters contributed further. The combined model had ROC AUC of 0.73 (p<0.001 LRT compared to the univariate mADC model).

Conclusion:

Entropy provides important additional information to mADC, thereby improving quantitative assessment of prostate MRI.

Limitations:

All patients were examined on a single scanner system, limiting the generalisability of our results, however, providing a highly standardised mADC measurement which entropy was able to provide added value to.

Ethics committee approval

Approved by an ethics committee with written informed consent waived (S-156/2018).

Funding:

No funding was received for this work.

Purpose:

To examine the added-value of DCE in multi-parametric prostate MRI with a region-of-interest based quantitative evaluation.

Methods and materials:

Clinical lesions reported in 3 Tesla MR exams from 315 consecutive patients with suspicion for prostate cancer were retrospectively segmented on ADC maps and a visually identified early DCE time-point. DCE was normalised to minimally enhancing parenchyma (DCEnorm). Multiple heuristic and pharmacokinetic parameters were determined, including the difference in bolus arrival time between the femoral artery and lesion (BATlesdiff) and normalised lesion AUC (AUCnorm). A basic logistic regression model included mean ADC and DCEnorm while a second model was extended by heuristic and pharmacokinetic parameters. The binary outcome was a Gleason grade group (GGG) >=2 from a targeted biopsy core for lesion-based analysis and maximum targeted GGG for patient-based analysis.

Results:

Of 308 patients with successful exams, 274 had 454 MR lesions that served as biopsy targets. 211 patients had a total of 275 MR lesions in the PZ. In the PZ, the model was simplified to ADCmean, DCEnorm (both p<0.001), BATlesdiff (p=0.04), and AUCnorm (p=0.15) by variable selection. This model performed significantly better (patient-based: AUROC 0.84 vs 0.80, p=0.04; lesion-based: AUROC 0.80 vs 0.76, p=0.03) than the ADC-only model. In the TZ, the model was reduced to the ADC-only model and no added benefit of any DCE-parameter was found.

Conclusion:

In a comprehensive quantitative ROI-based analysis, DCE demonstrates the ability to improve MRI assessment in the peripheral zone but not the transition zone.

Limitations:

Our results provide data on a typical PI-RADS compliant DCE protocol, however, cannot provide information on how to best trade off spatial for temporal resolution.

Ethics committee approval

Approved by an ethics committee with written informed consent waived (S-156/2018).

Funding:

No funding was received for this work.

Purpose:

To develop and validate a classifier system for the prediction of prostate cancer (PCa) aggressiveness using MRI texture analysis.

Methods and materials:

106 patients with histologically confirmed PCa were included in this retrospective study for model development and internal validation. Another 51 patients were included for independent external validation. A total of 64 first-order parameters and second-order texture parameters derived from the grey-level co-occurrence matrix were extracted from manually segmented tumours on T2w MRI and ADC maps. Data was analysed with a parametric test and Lasso logistic regression was used for feature selection and developing radiomics signatures to discriminate tumours with 4+3 vs > 3+4 GS high- and low-grade disease respectively (HG and LG). The predictive performance of the signature was evaluated via a receiver operating curve (ROC).

Results:

A total of 25 radiomics features significantly different between the two groups (highest correlation for ADC entropy: r=0.33, p<0.001) in the training group. We developed two radiomics signature based on 6 (l=min) and 1 (l=1.SE) features, respectively. The signatures resulted in AUC of 0.74 (0.65-0.84), p<0.001 and AUC of 0.73 (0.63-0.82), p<0.001, in the training group, and 0.64 (0.48-0.80), p<0.001 and AUC of 0.65 (0.49-0.80), p<0.001, in the validation set.

Conclusion:

MRI texture features could serve as potential diagnostic markers in assessing PCa biological aggressiveness.

Limitations:

A retrospective analysis, the absence of evaluation of contrast-enhanced sequences and maps, and relatively poor AUC values.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To assess to what extent radiomic features computed on high b-value DWI sequences (b=2000s/mm²) could reliably detect non-clinically significant (NCS) prostate cancer and reduce overtreatment.

Methods and materials:

This study retrospectively enrolled 25 patients of our institution, randomly extracted from PACS with a clinical suspicion of PCa who underwent prostate 3T-mpMRI. 10 patients reported NCS-PCa after TRUS biopsy, with a Gleason Score (GS)≤3+3, 15 were CS-PCa. PCa regions of interest (ROIs) were outlined in all slices by two experienced radiologists in consensus and reported on the DWI sequences, when needed, where 84 radiomic features with the corresponding ROC curves were computed. In order to prevent overfitting, a one-only feature was selected yielding the highest AUC and p-value<0.001 at the one-tail Wilcoxon rank-sum test.

Results:

The dispersion of local skewness (LS) of DWI values is higher for CS-PCa (p-value~10^-4) and AUC=0.92 (95%CI, 0.70-0.99). Sensitivity and specificity for NCS were 90% and 87%, respectively (1 FN and 2 FP), with a false omission rate (FOR) equal to 7%, this representing a very low risk of overtreatment. Moreover, the two FPs have GS=3+4, the CS-PCa group closest to NCS one.

Conclusion:

Radiomic features extracted from high b-values DWI sequences allows highlighting non-visible image properties related to the complexity of the tumour habitat. The higher variability of LS hints at increasing heterogeneity of tumour micro-environment for CS-PCa. In addition, this excellent performance stresses the promising role of DWI-based radiomics in discriminating CS-PCa from NCS-PCa.

Limitations:

No clinical parameters were considered for differentiation. However, at most, they could improve these results. In addition, the number of patients is limited, but uneven in their characteristics, since not derived from any dedicated study.

Ethics committee approval

IRB approval, written informed consent was waived.

Funding:

No funding was received for this work.

Purpose:

To assess the value of dynamic contrast-enhancement (DCE) use in detecting prostatic index lesions when evaluating the performance of two radiologists with different experience with biparametric (bpMRI) and multiparametric (mpMRI) examinations.

Methods and materials:

A retrospective study of 150 patients was performed including 3 Tesla prostate mpMRI.

Two radiologists, with advanced and limited experience, respectively, blinded to clinical and histological data, classified each index lesion using PI-RADS v2. Images were revisited with a reading, including diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) maps, T2-weighted (T2W) imaging, and, after 3 months, DCE.

Results were matched with Gleason patterns. The performance was quantified by sensitivity (SNS), specificity (SPC), and area under the curve (AUC) of the ROC (receiver operating characteristics).

Results:

Concordance was good for the expert reader: weighted Cohen’s k ≈ 0.809 (95% CI 0.707-0.912), unlike the performance of the inexperienced reader, which resulted in a Cohen's k 0.396 (95% CI 0.241-0.551). The expert reader performed as well in bpMRI as in mpMRI (SNS=0.73–0.70, AUC=0.744–0.819; p=0.087). The inexperienced reader performed well in mpMRI, but significantly worse in bpMR: SNS=27.94 versus 64.71 and AUC=0.634 versus 0.807 (p=0.0024).

Conclusion:

Results based on the expert reader showed no substantial differences between the performance with bpMRI and mpMRI. The experience gained in mpMRI could advantage the reader in the transition to bpMRI.

Quite differently, the outcome for the inexperienced reader in mpMRI was similar to the expert reader, but in the absence of DCE, the performance dropped significantly.

Limitations:

The involvement of one radiologist for each category (experienced and non-experienced) could be limiting in the evaluation of results.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

Radiomics applied to magnetic resonance imaging (MRI) has shown promising results in classifying prostate cancer (PCa). However, the effects of these models on unseen data from different centres have rarely been addressed. Our goal is to evaluate the generalisability of radiomic models in the context of PCa classification and to compare the performance between our models and radiologists using prostate imaging reporting and data system (PIRADS) v2.

Methods and materials:

The data comprised multiparametric MRI, histology of radical prostatectomy, and pathology reports of 107 patients from three centres. By correlating the MRI with histology, 204 lesions were identified. From each lesion, radiomics features were extracted. Radiomics models for discriminating between high-grade and low-grade lesions were automatically developed using machine learning, either single-centre through cross-validation, or multicenter. For comparison with the multicentre setting, a subset of the dataset was classified by two expert radiologists using PIRADS.

Results:

The single-centre models obtained a mean AUC of 0.75 for the internal cross-validation; when testing on unseen data, the mean AUC decreased to 0.54. In the multicentre setting, the radiologists obtained a mean AUC of 0.47, while the radiomics model obtained a mean AUC of 0.75.

Conclusion:

Radiomic models may obtain a decent performance when tested in a single-centre setting. However, there can be a considerable drop in classification performance when testing these models on data from different centres. On a multicenter dataset, our radiomics model outperformed PIRADS and may represent a more accurate alternative for malignancy prediction.

Limitations:

Despite developing a multicenter-multivendor study, our cohort size is limited.

Ethics committee approval

The Medical ethics review committee of Erasmus MC (NL32105.078.10). The national Dutch trials register (ID: NL2368).

Funding:

Partly financed by the Netherlands Organization for Scientific Research.

Purpose:

To measure K^trans and correlate it with a Gleason score (GS) and a PI-RADS score in patients with prostate cancer.

Methods and materials:

This retrospective study included patients with pathologically proven prostate cancer who had undergone clinically indicated 1.5 Tesla multi-parametric magnetic resonance imaging (MRI) examination. T2 weighted (T2w) images, diffusion-weighted images (DWI), and dynamic contrast-enhanced (DCE) sequences were obtained. A PI-RADS score was calculated for all tumour lesions. From a DCE-MRI dataset, K^trans was computed and compared between patients with clinically insignificant (GS≤6) and clinically significant (GS≥7) prostate cancer. The Spearman rank-order correlation coefficient (ρ) was used to assess the correlation strength between K^trans and GS and between K^trans and a PI-RADS score.

Results:

21 patients (age: 67±12 years; BMI: 26.63±4.04 Kg/m²) with a PSA of 7.91±3.01 were included in the study. 7 patients (33.3%) had clinically insignificant prostate cancer while 14 patients (66.7%) were diagnosed with clinically significant prostate cancer. The mean K^trans value was 0.42±0.20 min^-1 (range: 0.15–0.75). K^trans was significantly higher (0.52±0.14 min^-1) in clinically significant prostate cancer compared to clinical insignificant prostate cancer (0.23±0.15 min^-1; P=0.016). K^trans showed moderate significant correlation with GS (ρ=0.575, P=0.006), while it showed no significant correlation with PI-RADS (ρ=0.386, P=0.069).

Conclusion:

K^trans may discriminate between clinically insignificant and significant prostate cancer and show a moderate correlation with GS. This MP-MRI may serve as an imaging biomarker in prostate cancer.

Limitations:

Limitations of this study are the small sample size and the inclusion of only patients with proven prostate cancer, despite its design requiring a selected patient population to assess the correlation between K^trans and GS.

Ethics committee approval

Written informed consent obtained.

Funding:

No funding was received for this work.

Purpose:

The indiscriminate use of contrast-enhanced MRI for the detection of PCa has been questioned in the new PIRADS v2.1. Our purpose is to compare multiparametric MRI (mpMRI) and biparametric MRI (bpMRI) for prostatic cancer (PCa) detection using PIRADSv 2.1.

Methods and materials:

The mpMRI exams performed between June and October 2019 were considered and lesions with a PIRADS from 3 to 5 in the reports were independently reviewed by two radiologists having 3 (reader A) and 10 (reader B) years experience. Evaluation of T2, DWI, ADC (bpMRI), and DCE (mpMRI) was performed, indicating a partial score for each parameter of bpMRI and a final PIRADS score for every lesion, both with and without DCE. Inter-rater agreement (Cohen's Kappa) was calculated regarding PIRADS for bpMRI and mpMRI. Data analysis was performed for cases upgraded to PIRADS 4-5 after DCE.

Results:

47 prostate MRIs were considered for 61 total lesions (47 PZ, 14 TZ). Inter-rater agreement was Kappa=0.9 and 0.8 (very good) for assigned PIRADS, respectively, for bpMRI and mpMRI. Inter-rater agreement was only moderate (Kappa=0.5) for T2 score: 11/13 lesions were in PZ and therefore this had little impact.

23 PZ lesions with PIRADS 3 were upgraded to PIRADS 4 or 5 after DCE (the same lesions for both readers). Out of 23 upgraded lesions, 13 for reader A and 6 for reader B did not show any worrisome feature in DWI or ADC map, while in the other cases, either DWI or ADC was considered as “markedly abnormal”.

Conclusion:

DCE appears to have a useful role in staging lesions with PIRADSv2.1 only for PZ lesions with a PIRADS 3 score, independent of the DWI features.

Limitations:

A small series.

Ethics committee approval

Written informed consent obtained.

Funding:

No funding was received for this work.

Purpose:

To develop and externally validate a stepwise logistic regression model based on MRI radiomic features to predict different Gleason scores (GS) of prostate cancer (PCa).

Methods and materials:

A dataset was composed of 97 patients (105 PCas), enrolled in two institutions with different MRI scanners and protocols, who underwent robot-assisted radical-prostatectomy for PCa after mpMRI. Lesions were classified as low aggressive (LA) if their pathological Gleason score was ≤3+4 and high aggressive (HA) if it was ≥4+3. A previously developed computer-aided diagnosis (CAD) system able to detect PCas was applied to the dataset and the output of CAD was used as a PCa segmentation mask. From the segmented masks of ADC maps and T2w images, 59 first- and second-order radiomics features were extracted (i.e. skewness, kurtosis, and texture parameters from GLCM and GLRLM matrices). Patients from institution-A were used as training and testing dataset, while patients from institution-B as a validation set. A stepwise logistic regression model was created using 70% of lesions as a training set and 30% as a testing set, all randomly selected. Since the composition of a training and testing set might affect the performance of the model, this selection was repeated 5 times and the model reaching the highest accuracy in the testing set was chosen and applied to the validation set.

Results:

Accuracy of the training set was 96.6% (54/56), with a sensitivity in detecting HA PCas of 95.6% (22/23) and a specificity of 96.9% (32/33). The accuracy of the testing set was 83.3% (20/24), with a sensitivity of 60% (6/10) and a specificity of 100% (14/14). the accuracy of the validation set was 68% (17/25), with a sensitivity of 40% (4/10) and a specificity of 86.7% (13/15).

Conclusion:

MRI radiomic features are promising markers in discriminating PCa aggressiveness on the histopathological level.

Limitations:

Further confirmations are required by larger series.

Ethics committee approval

Approved by an ethical committee.

Funding:

No funding was received for this work.

Purpose:

To investigate the potential role of radiomic and deep learning features extracted from multiparametric MRI (mp­MRI) in predicting prostate cancer (PCa) aggressiveness, correlated with a Gleason score (GS). The aim of this study is to define a predictive model to distinguish low-­grade (GS <=3+4) from intermediate/high­grade (GS >=4+3) PCa.

Methods and materials:

Our population was composed of 50 peripheral zone PCa patients (PIRADS score 3-­5) who underwent a 1.5 T mp­MRI and freehand transperineal MRI/US fusion-guided targeted biopsy (62 lesions: 41 with GS<=3+4 and 21 with GS>=4+3). The lesions were segmented on T2w images and ADC maps. Two analyses were then carried out. (i) Radiomic features were computed, identifying the most discriminative set (signature). A support vector machine was trained on 52 cases (35 GS<=3+4 and 17 GS>=4+3) and tested on the remaining 10 (6 GS<=3+4 and 4 GS>=4+3). (ii) Three deep learning architectures were trained on the same dataset, augmented with rigid and non-rigid deformations, in order to limit overfitting.

Results:

The radiomic approach, using 6 features, allowed the correct classification of all cases (accuracy 100%, sensitivity 100%, and specificity 100%). The best performing DL model (based on VGG16) achieved accuracy 87.3%, sensitivity 97.2%, and specificity 74% on T2w images.

Conclusion:

Radiomic features can be useful in the assessment of lesion aggressiveness. The DL features need more investigations on a larger population and a more balanced distribution.

Limitations:

Our main limitation is the small number of patients.

Ethics committee approval

The study was approved by our Institutional Review Board.

Funding:

No funding was received for this work.