Diagnostic accuracy of AI for bpMRI screening of prostate cancer: a systematic review and meta-analysis

Abstract

The primary objective of this systematic literature review and meta-analysis is to evaluate the quality of prognostic models created for screening of prostate cancer (PCa). Methods: The systematic search of publications from January 2019 to September 2023 in the electronic databases ELibrary, PubMed, Google Scholar, Web of Science and Research Gate was used in accordance with the PRISMA protocol. Two authors independently assessed the need for inclusion or exclusion of the relevant studies Results: This meta-analysis included 21 studies. In total, 3,630 patients, of which 47% were with prostate cancer and 53% with benign prostate tumors. The average age of patients was 67.1 (mainly from 36 to 90 years). Eighty one percent (81%) of studies were based on T2-weighted imaging (T2-WI), 57% on diffusion-weighted imaging (DWI), and 76% on the apparent diffusion coefficients imaging (ADC). Forty three percent (43%) of studies were devoted to a malignancy formation in the transitional zone (TZ), 33% to the peripheral zone (PZ) of the prostate gland. Fifty two percent (52%) of authors conducted research on the entire organ, without dividing it into zones. The analysis showed that the researchers used machine learning (ML) algorithms: MLR (multiple logistic regression, in 76%), SVM (support vector machine, in 38%) and RF (random forest, in 24%). According to a meta-analysis of ROC-AUC assessment in 73 prognostic models described in the publications we studied, using methodological random effects, a final ROC-AUC value of 0.793 [95%CI 0.768; 0.818], I2 = 86.71%, p0.001. The most predictive models are based on T2-WI + ADC protocol: 0.860 [95%CI 0.813; 0.907], and those models that were created according to the “white box” principle (0.834 [95%CI 0.806; 0.861]). For comparison the values for “black box” are (0.733 [95%CI 0.695; 0.771]). Models using MRI and physiological features were slightly more accurate than the MRI parameters alone (0.869 [95% CI 0.844, 0.895] vs. 0.779 [95% CI 0.751, 0.807]). Model accuracy was virtually the same across PZ and/or TZ studies. Conclusion: The results reveal the most promising AI models. However, the clinical applicability may require more rigorous institutional validation and evaluation of efficacy in the prospective studies.