Diagnostic accuracy of artificial intelligence for the screening of prostate cancer in biparametric magnetic resonance imaging: a systematic review

Abstract

BACKGROUND: Based on the latest published data, 40,137 new cases of prostate cancer were reported in Russia in 2021, ranking second after lung cancer in men. Thus, prostate cancer is one of the most common malignant neoplasms in men. Accurate and timely detection of prostate cancer is important under the current conditions. AIM: This systematic review aimed to assess the quality of prediction models designed to detect prostate cancer during initial presentation. MATERIALS AND METHODS: A systematic search was performed in eLibrary.ru, PubMed, Google Scholar, Web of Science, and ResearchGate for relevant publications indexed from January 2019 to September 2023 in accordance with the PRISMA protocol. Two authors independently assessed the relevant studies for potential inclusion or exclusion. RESULTS: This systematic review meta-analysis included 21 studies. In total, data from 3,630 patients were analyzed, of which 47% had prostate cancer and 53% had benign prostate neoplasms. The mean age of the patients was 67.1 (36–90) years. In addition, 81% of the studies were based on T2-weighted imaging, 57% on diffusion-weighted imaging, and 76% on apparent diffusion coefficient. Moreover, 43% and 33% of the studies were dedicated to transition zone and prostate peripheral zone neoplasms, respectively, and 52% of the authors examined the whole prostate gland, without dividing it into zones. The most common machine-learning algorithms applied by the investigators were as follows: multiple logistic regression (76%), support vector machine (38%), and random forest (24%). Based on the meta-analysis performed for the receiver operating characteristic-area under the curve (ROC–AUC) assessment with random-effect approach in 73 prediction models described in the publications, the final ROC–AUC was 0.793 [95% CI 0.768–0.818], I2 = 86.71%, p 0.001. The most accurate prediction models were based on the T2-weighted imaging + apparent diffusion coefficients imaging protocol: 0.860 [95% CI 0.813–0.907], and models created according to the “white box” principle (0.834 [95% CI 0.806–0.861]) were more accurate than the “black box” ones (0.733 [95% CI 0.695–0.771]). The models using radiomics and clinical features were slightly more accurate than those using the radiomics parameters alone (0.869 [95% CI 0.844–0.895] vs. 0.779 [95% CI 0.751–0.807]). Model accuracy was nearly identical across transitional and/or peripheral zone studies. CONCLUSIONS: Artificial intelligence demonstrated promising results. However, the clinical applicability may require more intensive expert inspection in healthcare institutions and evaluation of efficacy in prospective studies.