Diagnostic performance of magnetic resonance imaging-based radiomics for detecting prostate cancer: A systematic review and meta-analysis

Abstract

BACKGROUND

Prostate cancer (PCa) is the second most common malignancy and the fifth leading cause of cancer death among men worldwide. Magnetic resonance imaging (MRI)-based radiomics has emerged as a promising tool for diagnosing PCa, but its true potential remains a subject of ongoing debate.

AIM

To evaluate the diagnostic performance of MRI-based radiomics for PCa detection and compare it with the Prostate Imaging Reporting and Data System (PI-RADS) score.

METHODS

A systematic search of EMBASE, Web of Science, and PubMed was conducted up to August 18, 2025. Pooled sensitivity, specificity, and area under the curve were calculated. Subgroup analyses were performed to evaluate heterogeneity. Additionally, the diagnostic accuracy of MRI-based radiomics was compared with that of the PI-RADS score. Methodological quality was evaluated via the Radiomics Quality Score.

RESULTS

This meta-analysis included 49 studies encompassing 10512 patients. MRI-based radiomics demonstrated a pooled sensitivity of 0.84 [95% confidence interval (CI): 0.80-0.87], specificity of 0.78 (95%CI: 0.72-0.84), and area under the curve of 0.88 (95%CI: 0.85-0.91). Deek’s funnel plot asymmetry test indicated no publication bias. Subgroup analyses revealed that multiparametric MRI-based radiomics is more effective in diagnosing clinically significant PCa and that 3D-based radiomics outperforms 2D approaches. In a head-to-head comparison, the MRI-radiomics model yielded a numerically greater pooled diagnostic value (0.85; 95%CI: 0.82-0.88) than did the PI-RADS score (0.71; 95%CI: 0.63-0.77) (P = 0.07). The mean Radiomics Quality Score was 18.2 (50.6% of the maximum score). All studies reported performing cut-off analyses, 22 studies (44.9%) addressed biological correlates, and all claimed code or data accessibility.

CONCLUSION

MRI-based radiomics is a reliable tool for detecting PCa, with 3D radiomic models showing greater effectiveness than 2D approaches in terms of sensitivity (0.85 vs 0.79). Radiomics also offers superior diagnostic accuracy for clinically significant PCa compared with the PI-RADS score, underscoring its potential in improving PCa diagnostics.

Key Words: Prostate cancer; Radiomics; Magnetic resonance imaging; Prostate Imaging Reporting and Data System; Meta-analysis

Core Tip: This meta-analysis demonstrates that magnetic resonance imaging-based radiomics is a reliable tool for detecting prostate cancer, with 3D radiomic models offering higher sensitivity than 2D models. Compared to the Prostate Imaging Reporting and Data System score, radiomics shows superior diagnostic accuracy for clinically significant prostate cancer, highlighting its potential to optimize diagnostic workflows and reduce unnecessary biopsies.

INTRODUCTION

Prostate cancer (PCa) is the most frequently diagnosed malignancy among men, with over 3 million cases reported in the United States alone. In 2025, an estimated 312000 new cases and 38000 deaths are projected, with rising global incidence attributed to aging populations[1]. Screening for PCa via prostate-specific antigen testing has demonstrated potential in reducing mortality rates. However, prostate-specific antigen testing also identifies indolent cancers that may not require treatment[2]. Multiparametric magnetic resonance imaging (mpMRI) has significantly advanced the detection and localization of primary PCa, as well as the evaluation of tumor aggressiveness. To standardize magnetic resonance imaging (MRI) interpretation and imaging protocols, the European Society of Urogenital Radiology and the American College of Radiology jointly introduced the Prostate Imaging Reporting and Data System (PI-RADS) in 2012, with an updated version 2.1 released in 2019[3]. While widely used in clinical settings, the PI-RADS faces challenges, including variability in interreader agreement, inconsistent diagnostic accuracy for PI-RADS category 3 lesions [5%-6% clinically significant PCa (csPCa) missed on targeted biopsies], and ambiguous clinical management pathways.

Radiomics is an emerging field in medicine that focuses on extracting quantitative feature information (known as radiomic features) from radiological images, which are often undetectable by the human eye, to develop clinical decision support systems[4,5]. By offering a noninvasive and reliable approach, radiomics models have the potential to enhance precision medicine, surpassing traditional models that rely solely on clinicopathological factors[6]. Recent studies have explored the application of radiomics in various aspects of PCa, including detection, prognosis, and treatment[7,8]. Radiomics analysis may also predict the likelihood of PCa in suspected cases prior to surgery, thereby facilitating more personalized treatment.

Despite advancements in radiomics for PCa diagnosis, several challenges remain that require further exploration: (1) The diagnostic potential of radiomics based on mpMRI, including dynamically contrast-enhanced sequences, compared with those excluding such sequences; (2) The efficacy of radiomics in detecting clinically significant cancers; (3) The comparative diagnostic performance of radiomics derived from 2D vs 3D images; and (4) The advantages of radiomics over the PI-RADS v2.1 system. Additionally, substantial variability in methodological pipelines across radiomic studies poses challenges to the generalizability and reproducibility of the findings. The Radiomic Quality Score (RQS) has recently been employed to assess key study attributes, including evidence strength, performance metrics, biological relationships, clinical validation, feature adjustment using tools like PyRadiomics (open-source platform), and repeatability[9]. Accordingly, the primary aim of this systematic review was to evaluate the literature on mpMRI-based radiomics for PCa detection. A secondary objective was to examine the methodological rigor of workflows used in radiomics research.

MATERIALS AND METHODS

This systematic review was prospectively registered with PROSPERO (CRD42022357840) and conducted in accordance with the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies guidelines. To ensure accuracy and reliability, two independent reviewers independently carried out each stage of the review process, including title and abstract screening, full-text screening, data extraction, assessment of adherence to reporting guidelines, bias, and applicability.

Search strategy

A comprehensive search was performed across three main medical literature databases - EMBASE, Web of Science, and PubMed - from inception until August 18, 2025. The search employed a straightforward query without database-specific Boolean search operators: (Prostate cancer OR prostatic neoplasm OR prostate cancer patient) AND (radiomic characteristics OR radiomics analysis OR texture). Additionally, references from the included studies were manually reviewed to identify any relevant research.

Two reviewers (5 years and 3 years of meta-analysis experience, respectively) independently screened articles. The process began with title and abstract screening to exclude studies that did not meet the criteria. Full-text reviews were then conducted on the remaining studies to finalize the selection. Disagreements were resolved through discussion and consensus. For data pooling, studies providing information on either an internal or external test set were assessed separately. When multiple models were presented in a study, the model with the highest predictive performance was selected.

Inclusion and exclusion criteria

Only publications on original research studies that employed prostate biopsy or postoperative pathology as the diagnostic ‘gold standard’ were considered for inclusion. Publications such as reviews, meta-analyses, case reports, conference abstracts, letters to the editor, comments, posters, technical reports, duplicate studies, non-English publications, and nonhuman studies were excluded. Studies were also omitted if they failed to assess model performance via the area under the receiver operating characteristic (ROC) curve. Figure 1 provides a detailed overview of the study selection process.

Figure 1 PRISMA flow diagram for study selection. The diagram illustrates the process of identification, screening, eligibility assessment, and inclusion of studies for the meta-analysis.

Data extraction

Two reviewers, with 5 years and 3 years of experience in meta-analysis research, independently extracted key data, including study design, patient demographics, segmentation software, lesion type (2D/3D), intensity normalization, MRI technique (field strength, sequences), diagnostic criteria, and accuracy metrics (true and false positives and negatives). In cases where the data were unclear or could not be extracted into 2 × 2 tables, the study authors were contacted for clarification whenever feasible. Any discrepancies between the reviewers were resolved through discussion and consensus.

Quality assessment

The quality of the 49 included studies was evaluated via the RQS, which has a maximum possible score of 36. The scores ranged from 12 to 24, with a mean score of 18.2 ± 3.15, representing 50.6% of the total possible score. The low overall score was primarily due to common shortcomings in study design, such as the prevalent lack of prospective validation, phantom studies, and robust external validation. Notably, while all studies (100%) fulfilled basic reporting standards by including discrimination statistics and cut-off analyses, and all declared code or data accessibility, the practical implementation of open science varied. Only 22 studies (44.9%) attempted to explore the biological correlates of radiomic features.

Statistical analysis

The extracted data were summarized as means for continuous variables and percentages for categorical variables. Using the collected 2 × 2 contingency tables, a hierarchical random effects model was applied to pool diagnostic performance metrics, including sensitivity, specificity, and the diagnostic odds ratio. Summary estimates for sensitivity and specificity, along with their 95% confidence intervals (CI), were calculated via a bivariate model that accounted for both inter- and intra-study heterogeneity. The hierarchical summary ROC model was employed to generate the summary ROC curve and compute the area under the curve (AUC).

Such heterogeneity was evaluated via Cochran’s Q test and the I² statistic. The risk of publication bias was assessed through a funnel plot and regression analysis. Subgroup analyses were conducted to identify sources of heterogeneity among studies, focusing on three main factors: (1) Scanning sequence [with or without dynamic contrast enhancement (DCE)]; (2) Detection of csPCa vs nonclinically significant PCa (non-csPCa); and (3) Radiomics derived from 2D vs 3D imaging.

For head-to-head comparisons, individual study results and summary estimates were plotted in ROC space with connecting lines. Differences in sensitivity and specificity between models were tested via the z test for paired data. Statistical analyses were performed via Stata (version 17.0) with the “midas” package and R (version 4.2.1) with the “metafor” package.

RESULTS

Study selection

The electronic literature search yielded 1049 records. After removing 721 duplicates, 328 records remained. Screening titles and abstracts led to the exclusion of 121 records, leaving 109 full-text articles for detailed review. Ultimately, 49 studies, encompassing a total of 10512 patients, were included in the qualitative analysis. The studies were published online from September 2013 to August 2024. The baseline characteristics of the included studies are summarized in Table 1.

Table 1 Key characteristics of the included studies.

Ref.	Year	Country	Number of patients	Center	Field strength (T)	Study design (Re/Pr)	Sequences	Segmentation software	Lesion segmentation	Code/data availability	Biological correlation analysis	Cut-off analyses conducted	Reference standard
Bertelli et al[4]	2021	Italy	112	Single	1.5T	Re	mpMRI without DCE	Manual segmentation	2D	Yes	No	Yes	TRUSGB and MRGB
Wu et al[5]	2019	Canada	90	Single	3.0T	Re	mpMRI with DCE	ImageJ (v.1.48, National Institutes of Health, Bethesda, MD)	2D	Yes	Yes	Yes	RP
Urakami et al[6]	2022	Japan	101	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	No	Yes	RP
Nketiah et al[7]	2021	Norway	96	Multicenter	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Jensen et al[8]	2019	Denmark	182	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	TRUSGB
Gong et al[23]	2022	China	489	Single	3.0T	Re	mpMRI without DCE	ITK-SNAP, version 3.4.0	3D	Yes	No	Yes	RP
Castillo T et al[13]	2021	Netherlands	644	Multicenter	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	STRUSGB
Castillo T et al[14]	2021	Netherlands	204	Multicenter	3.0T and 1.5T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	RP
Orczyk et al[15]	2019	United Kingdom	20	Multicenter	1.5T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Damascelli et al[50]	2021	Italy	102	Single	1.5T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Niu et al[51]	2018	China	184	Single	1.5T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	NR
Peng et al[52]	2021	China	194	Single	1.5T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Zhang et al[24]	2021	China	139	Single	3.0T	Re	mpMRI without DCE	Open source, ITK-SNAP	3D	Yes	Yes	Yes	RP
Han et al[25]	2021	China	176	Single	3.0T	Re	mpMRI with DCE	ITK-SNAP, Toolbox v3.6.0	3D	Yes	No	Yes	MRI-TRUS biopsy
Gong et al[26]	2020	China	489	Single	3.0T	Re	mpMRI without DCE	ITK-SNAP v.3.4.0	3D	Yes	Yes	Yes	RP
Cheng et al[27]	2023	China	226	Single	3.0T	Re	mpMRI with DCE	ITKSNAP software	3D	Yes	No	Yes	STRUSGB and MRGB
Min et al[28]	2019	China	280	Single	3.0T	Re	mpMRI without DCE	ITK-SNAP software	3D	Yes	Yes	Yes	RP
Li et al[29]	2020	China	381	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Bonekamp et al[10]	2018	Germany	316	Single	3.0T	Pr	mpMRI without DCE	The medical imaging toolkit	3D	Yes	No	Yes	RP
Zhang et al[30]	2022	China	142	Single	3.0T	Re	mpMRI with DCE	ITK-SNAP software	3D	Yes	Yes	Yes	STRUSGB and MRGB
Liu et al[31]	2021	United States	402	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	RP
Hou et al[32]	2020	China	263	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Bleker et al[11]	2021	Netherlands	206	Single	3.0T	Pr	mpMRI with DCE	Manual segmentation	2D	Yes	No	Yes	RP
Xiong et al[53]	2021	China	85	Single	1.5T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Yang et al[33]	2023	China	392	Single	3.0T	Re	mpMRI without DCE	ITK-SNAP software	3D	Yes	Yes	Yes	RP
Bleker et al[16]	2020	Netherlands	262	Multicenter	3.0T	Re	mpMRI without DCE	Manual segmentation	2D	Yes	No	Yes	STRUSGB and MRGB
Woźnicki et al[20]	2020	Germany	191	Single	3.0T	Re	mpMRI without DCE	Board-certified radiologist (D.N.)	3D	Yes	Yes	Yes	RP
Bevilacqua et al[34]	2021	Italy	76	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	RP
Donisi et al[21]	2021	Italy	299	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	2D	Yes	Yes	Yes	TRUSGB
Zhou et al[17]	2023	China	170	Multicenter	3.0T	Re	mpMRI without DCE	ITK-SNAP software	3D	Yes	No	Yes	TRUSGB
Chen et al[35]	2019	China	381	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	TRUSGB
Fehr et al[36]	2015	United States	147	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Roest et al[18]	2023	Netherlands	1513	Multicenter	3.0T and 1.5T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	RP
Algohary et al[37]	2018	United States	301	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Zhang et al[38]	2021	China	140	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	2D	Yes	No	Yes	TTSB
Isaksson et al[39]	2020	Italy	121	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Tanadini-Lang et al[40]	2018	Switzerland	47	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	No	Yes	RP
Bourbonne et al[41]	2020	United States	195	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
McGarry et al[12]	2022	United States	48	Single	3.0T	Pr	mpMRI with DCE	Manual segmentation	2D	Yes	No	Yes	RP
Penzias et al[42]	2018	United States	34	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	NR
Daniel et al[19]	2019	Austria	42	Multicenter	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	NR
Wang et al[22]	2017	China	54	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	TRUSGB
Jung et al[43]	2020	Korea	68	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	TRUSGB and RP
Stoyanova et al[44]	2016	United States	46	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	No	Yes	TRUSGB and RP
Schieda et al[45]	2021	Canada	76	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Ginsburg et al[46]	2017	United States	54	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	No	Yes	RP
Shiradkar et al[47]	2018	United States	120	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	RP
Zhang et al[48]	2019	China	140	Single	3.0T	Re	mpMRI with DCE	Manual segmentation	3D	Yes	No	Yes	TRUSGB and MRGB
Toivonen et al[49]	2019	Finland	72	Single	3.0T	Re	mpMRI without DCE	Manual segmentation	3D	Yes	Yes	Yes	TRUSGB and RP

Re: Retrospective cohort study; Pr: Prospective cohort study; mpMRI: Multiparametric magnetic resonance imaging; DCE: Dynamic contrast enhancement; TRUSGB: Transrectal ultrasound-guided biopsy; MRGB: Magnetic resonance imaging-guided biopsy; RP: Radical prostatectomy; STRUSGB: Systematic transrectal ultrasound-guided biopsy; MR-TRUS: Fusion of magnetic resonance and transrectal ultrasound images; TTSB: Transperineal template saturation biopsy; NR: Not reported.

Among the 49 studies, three utilized a prospective enrollment methodology[10-12], whereas eight involved data from multiple institutions[7,13-19]. Seven studies employed multiple MRI scanners[5,7,8,14,20-22]. With respect to MRI field strength, a 3-T scanner was the most frequently used (41/49 studies)[5-8,10-13,16,17,19-49], with two studies employing both 1.5-T and 3-T scanners[14,18]. In terms of lesion segmentation, seven studies adopted a 2D approach[4,5,11,12,16,21,38], whereas the remaining studies utilized 3D segmentation[6-8,10,13-15,17-20,22-37,39-53].

Heterogeneity and diagnostic accuracy

The included studies exhibited heterogeneity based on Cochran’s Q tests (Q = 699.30 for sensitivity and Q = 921.59 for specificity, both P > 0.05) and I² statistics (I² = 92.42% for sensitivity and I² = 94.25% for specificity) (Figure 2). The data distribution in the summary receiver operating characteristic plot displayed a “shoulder-arm” pattern, suggesting the potential influence of a threshold effect among the studies. Analysis via Stata revealed an overall AUC of 0.88 (95%CI: 0.85-0.91) (Figure 3). Radiomics demonstrated a pooled sensitivity of 0.84 (95%CI: 0.80-0.87) and specificity of 0.78 (95%CI: 0.72-0.84) (Figure 2). Furthermore, Deeks’ funnel plot indicated no evidence of publication bias (P = 0.27; Figure 4).

Figure 2 Coupled forest plot of eligible studies. This plot illustrates the estimated sensitivities and specificities of the studies included in the meta-analysis. The horizontal bars denote the 95% confidence intervals for each study, the diamond represents the pooled random-effects rate, and the vertical line represents the line of no effect. CI: Confidence interval.

Figure 3 Summary receiver operating characteristic curves with prediction and confidence contours. The numbers within circles correspond to the specific studies included in the bivariate model. SROC: Summary receiver operating characteristic; SENS: Sensitivity; SPEC: Specificity; AUC: Area under the curve.

Figure 4 Deeks’ funnel plot. A P value of 0.27 suggests no evidence of publication bias. ESS: Effective sample size.

Subgroup analysis

Subgroup analyses were conducted to explore potential sources of heterogeneity on the basis of three factors: (1) Scanning sequence (with or without DCE); (2) Detection of csPCa vs noncsPC; and (3) Radiomics derived from 2D vs 3D imaging. A subgroup analysis of studies using mpMRI with DCE sequences (20 studies) revealed a pooled sensitivity of 0.88 (95%CI: 0.84-0.91) and specificity of 0.83 (95%CI: 0.76-0.88). In comparison, mpMRI studies without DCE sequences (29 studies) reported a pooled sensitivity of 0.82 (95%CI: 0.77-0.87) and specificity of 0.73 (95%CI: 0.67-0.83). Seven studies utilizing 2D radiomics achieved a pooled sensitivity of 0.79 (95%CI: 0.71-0.83), a specificity of 0.76 (95%CI: 0.73-0.83), and an AUC of 0.78 (95%CI: 0.73-0.81). In contrast, forty-two 3D radiomic analyses reported a pooled sensitivity of 0.85 (95%CI: 0.82-0.88), a specificity of 0.86 (95%CI: 0.77-0.89), and an AUC of 0.86 (95%CI: 0.82-0.95).

For the diagnosis of csPCa (30 studies), the pooled sensitivity was 0.85 (95%CI: 0.82-0.88), the specificity was 0.87 (95%CI: 0.84-0.92), and the AUC was 0.93 (95%CI: 0.81-0.95). For non-csPCa detection (19 studies), the pooled sensitivity was 0.71 (95%CI: 0.63-0.77), the specificity was 0.78 (95%CI: 0.73-0.82), and the AUC was 0.76 (95%CI: 0.72-0.83). The detailed results of these subgroup analyses are presented in Table 2.

Table 2 Results of the subgroup analysis, 95% confidence intervals.

	Number of studies	Sensitivity			P value	Specificity			P value	AUC
		Sensitivity	P value	I2		Specificity	P value	I2
Overall	49	0.84 (0.81-0.87)	< 0.01	93.08 (91.76-94.40)		0.78 (0.70-0.84)	< 0.01	94.68 (93.75-95.62)		0.88 (0.85-0.91)
Sequences					< 0.01				< 0.01
mpMRI with DCE	20	0.88 (0.84-0.91)	< 0.01	72.75 (60.62-84.87)		0.83 (0.76-0.88)	< 0.01	84.66 (78.82-90.50)		0.82 (0.81-0.97)
mpMRI without DCE	29	0.82 (0.77-0.87)	< 0.01	94.41 (93.11-95.71)		0.73 (0.67-0.83)	< 0.01	95.90 (95.03-96.77)		0.87 (0.81-0.94)
Lesion segmentation					< 0.01				< 0.01
2D	7	0.79 (0.71-0.83)	< 0.01	92.54 (88.50-96.59)		0.76 (0.73-0.83)	< 0.01	93.33 (89.82-96.84)		0.78 (0.73-0.81)
3D	42	0.85 (0.82-0.88)	< 0.01	92.21 (91.43-94.38)		0.86 (0.77-0.89)	< 0.01	94.56 (93.52-95.60)		0.86 (0.82-0.95)
Tumor aggressiveness					< 0.01				< 0.01
csPCa	30	0.85 (0.82-0.88)	< 0.01	95.02 (93.92-96.12)		0.87 (0.84-0.92)	< 0.01	96.21 (95.44-96.96)		0.93 (0.81-0.95)
Non-csPCa	19	0.71 (0.63-0.77)	< 0.01	69.89 (55.77-84.00)		0.78 (0.73-0.82)	< 0.01	83.84 (77.44-90.25)		0.76 (0.72-0.83)

AUC: Area under the curve; mpMRI: Multiparametric magnetic resonance imaging; DCE: Dynamic contrast enhancement; csPCa: Clinically significant prostate cancer; Non-csPCa: Nonclinically significant prostate cancer.

Head-to-head comparison

Sixteen studies conducted direct comparisons between MR-radiomics and PI-RADS for the diagnosis of PCa. The pooled AUC for MRI-radiomics (0.85; 95%CI: 0.82-0.88) was notably greater than that of PI-RADS (0.71; 95%CI: 0.63-0.77) (Figure 5). Although this difference approached statistical significance, it was not statistically significant (P = 0.07).

Figure 5 Meta-analysis of head-to-head comparisons. The plot displays the sensitivity and specificity in the receiver operating characteristic space for head-to-head comparisons between magnetic resonance imaging-radiomics and Prostate Imaging Reporting and Data System. The dotted lines represent the findings of the meta-analysis. MRI: Magnetic resonance imaging; PI-RADS: Prostate Imaging Reporting and Data System.

Quality assessment

The quality of the 49 included studies was evaluated via the RQS, which has a maximum possible score of 36. The scores ranged from 12 to 24, with a mean score of 18.2 ± 3.15, representing 50.6% of the total possible score (Figure 6). The relatively low RQS was attributed primarily to the lack of prospective designs, phantom studies, and validation in external cohorts. All studies reported discrimination statistics and performed cut-off analyses. Code or data were publicly accessible in 49 studies (100%), and 22 studies (44.9%) identified and addressed biological correlates.

Figure 6 Group bar plot of the Radiomic Quality Score by study. This plot shows the distribution of methodological quality scores (Radiomic Quality Score, maximum 36) across the 49 included.

DISCUSSION

This study presents a systematic review and meta-analysis of published articles investigating mpMRI combined with radiomic analysis for detecting PCa. A total of 49 studies involving 10512 patients met the inclusion criteria. The findings demonstrate that the mpMRI-based radiomic approach achieves high diagnostic performance, with a pooled sensitivity of 0.84 and specificity of 0.78. Subgroup analysis revealed that mpMRI-based radiomics is more sensitive in detecting csPCa than in detecting non-csPCa. Additionally, radiomics derived from the 3D volume of lesions provided a more comprehensive depiction of tumor characteristics than did radiomics based on the largest cross-sectional area. In head-to-head comparisons, mpMRI-based radiomics outperformed the traditional PI-RADS system, highlighting its potential superiority in clinical applications. However, quality assessment via the RQS yielded an average score of 18.2 for all the included studies, reflecting limitations in study design and methodology.

Compared with previous studies, this study has several advantages. The primary finding is that mpMRI-based radiomics is highly effective in diagnosing PCa. Notably, many researchers have suggested abandoning DCE-MRI because of its lengthy acquisition time, potential risks, additional costs, patient discomfort from contrast agent injections, and relatively limited incremental benefits. In a head-to-head comparison, mpMRI demonstrated significantly greater pooled sensitivity than bpMRI for PCa detection. Urakami et al[6] highlighted that combined histograms of DCE images could reveal temporally changing perfusion patterns within tumors, enabling the identification of abnormal microvascularization. This approach may facilitate stratification of PCa patients into low- and high-tumor groups, potentially aiding in personalized treatment strategies. Discerning csPCa from benign conditions or non-csPCa is vital, as it directly impacts patient management strategies[13,14]. Our findings indicate that the mpMRI-based radiomic model demonstrates substantially greater sensitivity and specificity for identifying csPCa than non-csPCa. Numerous studies have explored the potential of radiomic features for predicting PCa and csPCa[50,51]. For example, Niu et al[51] reported significant differences in texture features derived from T2W images and apparent diffusion coefficient maps when comparing benign tissue to PCa, as well as between high- and low-grade PCa. The large sample size of this study (10512 patients) strengthens the evidence supporting mpMRI-based radiomics as a reliable tool for detecting csPCa. By accurately identifying csPCa, this approach could enable clinicians to design tailored treatment plans and reduce the number of unnecessary biopsies, ultimately improving patient outcomes[24,52].

Manual region of interest drawing remains the most widely used segmentation approach in radiomics. Segmentation represents a critical first step in image processing, as it significantly influences the accuracy of radiomic assessment. Studies by Bleker et al[11] and Algohary et al[37] suggest that whole-tumor volumetric analysis provides a more comprehensive representation of tumor heterogeneity than cross-sectional analysis does. In this study, we explored two segmentation patterns - 2D- and 3D-based radiomics - for PCa diagnosis. The results indicate that 3D segmentation produces imaging histology features that are more effective for diagnosing PCa than those derived from 2D segmentation. The advantage may stem from the inability of a single cross-sectional slice to capture the most representative portions of tumor tissue, given the inherent diversity and heterogeneity of tumor imaging features across different levels[34]. By incorporating phenotypic variables for the entire prostate, the 3D model offers a more holistic characterization than the 2D model does.

While the introduction of PI-RADS has significantly increased the utility of prostate MRI, interreader variability remains a persistent challenge. To address this, computer-aided diagnostic approaches have gained traction in recent years[18,35,36]. Numerous studies have demonstrated that radiomics models outperform PI-RADS and clinical indicators in the diagnosis and assessment of the aggressiveness of PCa. However, few studies have directly compared the performance of radiomics and PI-RADS using mpMRI in a head-to-head manner. Our findings reveal that radiomics achieves superior diagnostic performance to that of PI-RADS when mpMRI is employed. These findings suggest that radiomics can augment the diagnostic capabilities of PI-RADS v2.1, offering a robust tool to enhance PCa diagnosis and guide clinical decision making.

The quality of radiomic studies is a critical factor in driving the advancement of radiomic research and its integration into future clinical practice. To evaluate the reliability and reproducibility of the radiomic findings, this study utilized the RQS, a robust tool introduced by Lambin et al[54], to assess the quality of radiomic investigations. The RQS results from this meta-analysis highlighted significant variability in the quality and methodologies of the included studies. These findings underscore the need for future radiomics research to adopt more rigorous methodologies and transparent reporting practices to improve the reliability and reproducibility of MRI-based radiomic models for PCa prediction. While all included studies reported performing cut-off analyses and claimed code/data accessibility, the actual public availability and reusability of these resources in practice require further scrutiny. Slightly less than half of the studies (44.9%) attempted to explore the biological relevance of radiomic features, a gap that reduces biological interpretability and impedes widespread clinical application.

This study had several limitations. First, the variability in scanner types, imaging protocols and tumor detection criteria across studies may have influenced the accuracy of the results. Second, the relatively small number of studies included in the quantitative synthesis, coupled with significant (but expected) heterogeneity, necessitates cautious interpretation of the pooled data. However, such heterogeneity is a common challenge in diagnostic accuracy meta-analyses, and we addressed this by investigating its origins through subgroup analyses. Finally, this review did not include a search of the gray literature. While this means that some important studies may have been overlooked, gray literature searches lack standardization, and the reliability of such sources can be uncertain. Additionally, 46 out of the 49 included studies (approximately 94%) were retrospective in design. Retrospective studies are susceptible to selection bias and variations in patient enrollment criteria, which may affect the generalizability and reliability of our pooled results. Moreover, the predominance of retrospective studies in our analysis calls for prospective validation to confirm these findings.

CONCLUSION

Radiomics analysis holds significant potential in addressing some of the current challenges in PCa diagnosis. Subgroup analyses revealed that a 3D volumetric approach offers superior sensitivity for PCa detection compared to 2D approaches. Furthermore, MRI-based radiomics showed higher diagnostic accuracy for csPCa than for non-csPCa. In a head-to-head comparison, the mpMRI-based radiomic method surpassed the PI-RADS score in diagnostic accuracy. However, future studies should aim for higher methodological rigor as measured by tools like the RQS. Efforts should focus on conducting prospective validations, ensuring truly accessible and reproducible code/data, and, crucially, investigating the biological underpinnings of radiomic features to improve clinical translation.

ACKNOWLEDGEMENTS

The authors thank the researchers and study participants whose work was included in this meta-analysis. We also appreciate the valuable suggestions provided by our colleagues during the preparation of this review.