Development of a prognostic model for diagnosis of prostate cancer based on radiomics of biparametric magnetic resonance imaging apparent diffusion coefficient maps and stacking of machine learning algorithms
- Authors: Kuznetsov A.I.1
-
Affiliations:
- Moscow Aviation Institute (National Research University)
- Issue: Vol 5, No 1S (2024)
- Pages: 80-82
- Section: Articles by YOUNG SCIENTISTS
- URL: https://journals.rcsi.science/DD/article/view/261112
- DOI: https://doi.org/10.17816/DD626145
- ID: 261112
Cite item
Full Text
Abstract
BACKGROUND: Prostate cancer is one of the most common cancers among men [1, 2]. In recent years, a number of prognostic models based on texture analysis of biparametric magnetic resonance images have been created. The research has shown that radiomics features extracted from apparent diffusion coefficient maps are the most reproducible [3]. However, the models were limited in accuracy, since they are built using a single machine learning algorithm, which takes into account only linear dependences [4–6].
AIM: Increasing the accuracy of a prognostic model diagnosing prostate cancer through the use of stacking machine learning algorithms that takes into account not only linear, but also nonlinear dependencies based on radiomics of biparametric magnetic resonance imaging apparent diffusion coefficient maps.
MATERIALS AND METHODS: A single-center cohort retrospective study of patients with suspected prostate cancer was conducted in the X-ray Diagnostics and Tomography Department of the United Hospital and Polyclinic (Moscow, Russia) from 2017 to 2023. The presence of prostate cancer was confirmed by biopsy or radical prostatectomy. Statistical analyses was performed using Python 3.11.
RESULTS: The study involved 67 men aged 60 [54; 66] years, of which 57 were diagnosed with prostate cancer, and 10 — with benign prostate formation. The LIFEx software identified 96 radiomic features.
Statistically significant differences were found for: PARAMS_ZSpatialResampling (the voxel size of the image: Z dimension) (p=0.001), SHAPE_Sphericity[onlyFor3DROI] (how spherical a Volume of Interest is) (p=0.006), SHAPE_Compacity[onlyFor3DROI] (how compact the Volume of Interest is) (p=0.004), GLRLM_HGRE (p=0.039), GLRLM_SRHGE (p=0.041), GLRLM_RLNU (p=0.039), where GLRLM — Grey-Level Run Length Matrix. Univariate logistic regression showed that SHAPE_Compacity[onlyFor3DROI] (R2=15%) and PARAMS_ZSpatialResampling (R2=18%) had a statistically significant effect on the outcome. First, using the multivariate logistic regression method, a prognostic model was built that takes into account only linear dependencies. The model includes 3 features that together have a statistically significant effect on the outcome (R2=23%): SHAPE_Sphericity[onlyFor3DROI], PARAMS_ZSpatialResampling and GLRLM_RLNU.
To describe nonlinear relationships, another model was built based on the “Decision Tree” algorithm. It included 4 indicators (R2=58%): DISCRETIZED_HISTO_Entropy_log10 (the randomness of the distribution), SHAPE_Sphericity[onlyFor3DROI], PARAMS_ZSpatialResampling and GLRLM_SRE.
Stacking of algorithms, which consists of calculating the arithmetic mean between the predictions of the multivariate logistic regression and “Decision Tree” algorithms, made it possible to construct a model that takes into account both linear and nonlinear dependencies. The model includes 5 features (R2=77%). The constructed model formed the basis of the developed calculator program [7], currently introduced into a radiology practice.
CONCLUSION: The new model built on the basis of apparent diffusion coefficient maps performs better (area under ROC-curve 99.0% [97.7; 100.0]) than the existing models with area under ROC-curve 83.6% [78.3; 88.9], which also show high heterogeneity (I2=71%). The accuracy of the new model was increased due to the use of stacking machine learning technologies, which made it possible to take into account both linear and nonlinear effects from features on the outcome.
Full Text
##article.viewOnOriginalSite##About the authors
Anton I. Kuznetsov
Moscow Aviation Institute (National Research University)
Author for correspondence.
Email: drednout5786@yandex.ru
ORCID iD: 0000-0003-2182-5792
SPIN-code: 8824-9080
Russian Federation, Moscow
References
- Mottet N, van den Bergh RCN, Briers E, et al. EAU-EANM-ESTRO-ESUR-SIOG Guidelines on Prostate Cancer—2020 Update. Part 1: Screening, Diagnosis, and Local Treatment with Curative Intent. Eur Urol. 2021;79(2):243–262. doi: 10.1016/j.eururo.2020.09.042
- Smelov PA, Nikitina SYu, editors. Health care in Russia. 2021: statistical compendium. Moscow: Federal'naya sluzhba gosudarstvennoi statistiki; 2021. (In Russ).
- Shah V, Turkbey B, Mani H, et al. Decision support system for localizing prostate cancer based on multiparametric magnetic resonance imaging. Med Phys. 2012;39(7):4093–4103. doi: 10.1118/1.4722753
- He D, Wang X, Fu C, et al. MRI-based radiomics models to assess prostate cancer, extracapsular extension and positive surgical margins. Cancer Imaging. 2021;21(1):46. doi: 10.1186/s40644-021-00414-6
- Lu Y, Li B, Huang H, et al. Biparametric MRI-based radiomics classifiers for the detection of prostate cancer in patients with PSA serum levels of 4~10 ng/mL. Front Oncol. 2022;12. doi: 10.3389/fonc.2022.1020317
- Chen T, Li M, Gu Y, et al. Prostate Cancer Differentiation and Aggressiveness: Assessment With a Radiomic-Based Model vs. PI-RADS v2. Journal of Magnetic Resonance Imaging. 2019;49(3):875–884. doi: 10.1002/jmri.26243
- Certificate for the computer software 2023669718/ 19.09.2023. Schepkina EV, Kruchkova OV, Kuznetsov AI. Computer Software for the Prognosis of Probability of Prostate Cancer in Patients Based on MRI. Available from: https://www.elibrary.ru/item.asp?id=54657407 [cited 2024 Jan 28]. (In Russ). EDN: FNYFEB
![](/img/style/loading.gif)