0859

Multiparametric MRI Radiomic Features Improve Patient Selection for Active Surveillance in Prostate Cancer
Veronica Wallaengen1, Evangelia I Zacharaki1, Mohammed Alhusseini1, Isabella M Kimbel1, Nachiketh Soodana Prakash2, Ahmad Algohary1, Adrian L Breto1, Sandra M Gaston1, Rosa P Castillo Acosta3, Oleksandr N Kryvenko4, Bruno Nahar2, Dipen J Parekh2, Alan Pollack1, Sanoj Punnen2, and Radka Stoyanova1
1Department of Radiation Oncology, University of Miami Miller School of Medicine, Miami, FL, United States, 2Department of Urology, University of Miami Miller School of Medicine, Miami, FL, United States, 3Department of Radiology, University of Miami Miller School of Medicine, Miami, FL, United States, 4Department of Pathology, University of Miami Miller School of Medicine, Miami, FL, United States

Synopsis

Keywords: Diagnosis/Prediction, Cancer

Motivation: Accurate selection of prostate cancer patients to undergo active surveillance (AS) is crucial to ensure suitable treatment.

Goal(s): To develop an automated framework for mpMRI analysis to assist clinical decision making about whether a patient should remain on AS.

Approach: We developed a progression risk stratification model using mpMRI data from an AS trial, and incorporating clinical biomarkers and radiomic features from lesions identified by a deep neural network.

Results: The lesion segmentation network achieved a median DSC of 60.7%, and the progression prediction model an AUC of 81.1% in determining likelihood of progression within 12 months.

Impact: We present a fully automated methodology to assess prostate cancer progression risk for AS patients within the timeframe between their follow-up visits, thereby providing essential data for clinicians that can prospectively improve AS patient selection.

Introduction

Out of the around 250,000 patients who are yearly diagnosed with prostate cancer in the U.S, about half are at low risk1 and may consider active surveillance (AS)2,3, a safe alternative to immediate treatment that can reduce the burden of overtreatment. Yet, due to a lack of robust risk-stratification tools predicting histological progression, uncertainties around optimal patient selection remain. Identification of suitable AS candidates is of primary importance to (i) avoid delays in treatment for patients with high-risk of progression as emerging data show an increased risk of metastasis with long follow-up4, (ii) postpone treatment to preserve quality of life for low-risk patients, (iii) safely tailor the intensity of follow-up biopsies. Guidelines for reading prostate multiparametric MRI (mpMRI) using the standardized Prostate Imaging Reporting and Data System (PI-RADS)5, follow a semi-quantitative assessment, susceptible to low inter-reader agreement (<50%) and sub-optimal interpretation6,7. Among the many implemented computer-aided techniques for quantitative mpMRI analysis8,9, methods predicting histological progression are limited10,11. Based on recent findings10, our hypothesis is that using mpMRI to uncover the problem of tumor heterogeneity through quantitative imaging combined with clinical biomarkers can improve risk estimation, resulting in reliable identification of AS candidates.

Methods

We present a method to predict likelihood of change in prostate cancer histology that consists of three integrated steps (Figure 1): (i) training a neural network by deep learning (DL) for automatic segmentation of prostate and lesions suspicious for cancer; (ii) application of the network to identify lesions on mpMRI for AS patients; (iii) image intensity normalization; and (iv) extraction of radiomics features from the lesions, (v) predicting the likelihood of progression by fusing radiomics with clinical variables.Data from two patient cohorts were utilized for development of the lesion segmentation and risk prediction model, respectively (Table 1):
  • RP cohort (n=45): mpMRI data with lesions annotated after radical prostatectomy (RP)12.
  • MAST cohort (n=151): confirmatory mpMRI and clinical data from AS patients enrolled in the MAST trial (Figure 2). The data were further divided into: (a) Training and validation (n=135): patients who showed histological progression at the confirmatory visit or at the first follow-up visit (n=67), and patients who completed the full study without progressing or who are currently on AS with at least the first year follow-up completed without progressing (n=68). (b) Test (n=16): patients who progressed after 12 months were manually reviewed and the time of progression was retroactively adjusted to between 0-24 months and used as ground truth to verify the progression risk model (adjustment applied to 37.5% of patients).
A 3D nnU-Net13 architecture was used to build prostate and lesion segmentation networks, trained using 5-fold cross-validation with four mpMRI modalities from the RP cohort: T2-weighted (T2W), high B-value (BVAL), apparent diffusion coefficient (ADC), and early enhancing series from the dynamic contrast enhanced (DCE) sequence. The trained models were subsequently applied to segment prostate and lesions on the MAST cohort. After T2W signal intensity normalization14, a series of mpMRI histogram descriptors were extracted from the lesion volumes in the MAST cohort training and validation subset through a comprehensive radiomics pipeline. The number of features was reduced through feature selection and incorporated into a logistic regression model along with two clinical biomarkers indicative of progression: maximum PI-RADS score and Prostate-specific antigen (PSA) level (Figure 3). The final clinical-radiomics fused model was trained to discriminate between patients with low and high risk of histological progression within the timeframe of their next follow-up visit (12 months).

Results

Lesion detection, assessed using the Dice Similarity Coefficient (DSC) through criteria defined in recent work of others15,16, resulted in a median DSC of 0.607 (0.130-0.958). Figure 4a shows the predicted prostate and lesion masks compared to manual masks. The clinical-radiomics model achieved a mean AUC of 0.811±0.01 (averaged over 100 repetitions) on the validation set (Figure 4b). The sensitivity of the model in correctly labeling patients as high-risk was 66.7% when applied to an unseen test set (MAST test subset) (Figure 4c).

Discussion and Conclusions

The median DSC of the lesion segmentation network falls in the upper bound compared to similar studies15. The prediction sensitivity among missed progressors indicates that undetected progression, accumulated over 6 years, can be avoided by using the progression risk model. The model performance is promising for future clinical utilization to identify low-risk patients who will benefit by avoiding biopsies and remain on AS using monitoring through imaging only. More importantly, the model can help in avoiding delays in treatment of patients identified as high-risk.

Acknowledgements

The research was supported by the National Cancer Institute of the National Institutes of Health under Award Number P30CA240139, RO1CA189295, R01CA190105, and U01CA239141.

References

1. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin 2023;73(1):17-48.

2. Schaeffer E, Srinivas S, Antonarakis ES, Armstrong AJ, Bekelman JE, Cheng H, et al. NCCN guidelines insights: prostate cancer, version 1.2021. J Natl Compr Canc Netw. 2021;19:134–43.

3. Dall’Era MA, Albertsen PC, Bangma C, Carroll PR, Carter HB, Cooperberg MR, Freedland SJ, Klotz LH, Parker C, Soloway MS. Active surveillance for prostate cancer: a systematic review of the literature. European urology. 2012 Dec 1;62(6):976-83.

4. Yamamoto, T. et al. Metastatic Prostate Cancer in Men Initially Treated with Active Surveillance. The Journal of urology 195, 1409-1414, doi:10.1016/j.juro.2015.11.075 (2016).

5. Barentsz JO, Richenberg J, Clements R, et al. ESUR prostate MR guidelines 2012. Eur Radiol. 2012;22(4):746-757.

6. Rosenkrantz AB, Ginocchio LA, Cornfeld D, Froemming AT, Gupta RT, Turkbey B, Westphalen AC, Babb JS, Margolis DJ. Interobserver reproducibility of the PI-RADS version 2 lexicon: a multicenter study of six experienced prostate radiologists. Radiology. 2016 Sep;280(3):793-804.

7. Westphalen AC, McCulloch CE, Anaokar JM, Arora S, Barashi NS, Barentsz JO, Bathala TK, Bittencourt LK, Booker MT, Braxton VG, Carroll PR. Variability of the positive predictive value of PI-RADS for prostate MRI across 26 centers: experience of the society of abdominal radiology prostate cancer disease-focused panel. Radiology. 2020 Jul;296(1):76-84.

8. Delgadillo R, Ford JC, Abramowitz MC, Dal Pra A, Pollack A, Stoyanova R. The role of radiomics in prostate cancer radiotherapy. Strahlenther Onkol. 2020;196(10):900-912.

9. Lemaître G, Martí R, Freixenet J, Vilanova JC, Walker PM, Meriaudeau F. Computer-aided detection and diagnosis for prostate cancer based on mono and multi-parametric MRI: a review. Computers in biology and medicine. 2015 May 1;60:8-31.

10. Ploussard G, Rouvière O, Rouprêt M, van den Bergh R, Renard-Penna R. The current role of MRI for guiding active surveillance in prostate cancer. Nature Reviews Urology. 2022 Jun;19(6):357-65.

11. Rajwa P, Pradere B, Quhal F, Mori K, Laukhtina E, Huebner NA, D’andrea D, Krzywon A, Shim SR, Baltzer PA, Renard-Penna R. Reliability of serial prostate magnetic resonance imaging to detect prostate cancer progression during active surveillance: a systematic review and meta-analysis. European urology. 2021 Nov 1;80(5):549-63.

12. Tschudi Y, Pollack A, Punnen S, et al. Automatic Detection of Prostate Tumor Habitats using Diffusion MRI. Sci Rep. 2018;8(1):16801.

13. Isensee F. et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18, 203-211, doi: 10.1038/s41592-020-01008-z (2020).

14. Algohary A, Zacharaki EI, Breto AL, Alhusseini M, Wallaengen V, Xu IR, Gaston SM, Punnen S, Castillo P, Pattany PM, Kryvenko ON, Spieler B, Abramowitz MC, Dal Pra A, Ford JC, Pollack A, Stoyanova R. Uncovering Prostate Cancer Aggressiveness Signal in T2-weighted MRI through a Deep Learning-based Normalization Technique, NMR in Biomedicine, in press.

15. Simeth J, Jiang J, Nosov A, Wibmer A, Zelefsky M, Tyagi N, Veeraraghavan H. Deep learning‐based dominant index lesion segmentation for MR‐guided radiation therapy of prostate cancer. Medical Physics. 2023 Mar 1.

16. Saha A, Hosseinzadeh M, Huisman H. End-to-end prostate cancer detection in bpMRI via 3D CNNs: effects of attention mechanisms, clinical priori and decoupled false positive reduction. Medical image analysis. 2021 Oct 1;73:102155.

17. Peng H, Long F, Ding C. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1226-1238, Aug. 2005, doi: 10.1109/TPAMI.2005.159.

Figures

Figure 1. Overview of the risk stratification approach in clinical management of prostate cancer and implementation of the pipeline.

Table 1. Patient characteristics for study cohorts. All clinical biomarkers measured at initial confirmatory visit.

Figure 2. MAST (MRI-Guided Active Selection for Treatment of Prostate Cancer) trial overview (ClinicalTrials.gov: NCT02242773). Participants underwent mpMRI and MRI-US fusion biopsy within 18 months of diagnosis and every subsequent year until the 3rd year of the trial. Patients remained under AS as long as they did not show histological progression or decided to undergo treatment for other reasons.

Figure 3. Radiomics pipeline. T2W intensities normalized using a multireference tissue normalization approach14, BVAL and DCE normalized by gluteus maximus. 11 intensity features were extracted from each modality and the prostate volume, combined lesion volume, and number of lesions were added as features. The mRMR17 method was applied to select the most significant features. An exhaustive feature search was then performed for a logistic regression model with respect to optimizing the Area Under the Receiver Operating Characteristic Curve (AUC) for prediction of progression.

Figure 4. (a) Prostate masks (magenta) and lesion masks (green) manually outlined by radiologist (upper) and generated by the nnU-Net (lower) superimposed on three axial T2W slices for one of the patients in the MAST cohort, (b) ROC-AUCs of the 5 stratified folds (80%/20% train/test split) of the clinical-radiomic logistic regression model predicting histological progression (n=135), (c) Performance metrics of progression risk model when applied to test set (n=16).

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)
0859
DOI: https://doi.org/10.58530/2024/0859