Prostate MRI Using Deep Learning Reconstruction in Response to Cancer Screening Demands—A Systematic Review and Meta-Analysis

Abstract

Background/Objectives: There is a growing need for efficient prostate MRI protocols due to their increasing use in managing prostate cancer (PCa) and potential inclusion in screening. Deep learning reconstruction (DLR) may enhance MR acquisitions and improve image quality compared to conventional acceleration techniques. This systematic review examines DLR approaches to prostate MRI. Methods: A search of PubMed, Web of Science, and Google Scholar identified eligible studies comparing DLR to conventional reconstruction for prostate imaging. A narrative synthesis was performed to summarize the impact of DLR on acquisition time, image quality, and diagnostic performance. Results: Thirty-three studies showed that DLR can reduce acquisition times for T₂w and DWI imaging while maintaining or improving image quality. It did not significantly affect clinical tasks, such as biopsy decisions, and performed comparably to human readers in PI-RADS scoring and the detection of extraprostatic extension. However, AI models trained on conventional data might be less accurate with DLR images. The heterogeneity in image quality metrics among the studies prevented quantitative synthesis. Discussion: DLR has the potential to achieve substantial time savings in prostate MRI while maintaining image quality, which is especially relevant because of increased MRI demands. Future research should address the effect of DLR on clinically relevant downstream tasks, including AI algorithms’ performances and biopsy decisions, and explore task-specific accelerated protocols for screening, image-guided biopsy, and treatment.

1. Introduction

The demand for magnetic resonance imaging (MRI) is increasing as its role in prostate cancer (PCa) early diagnosis, staging, active surveillance, follow-up, radiation treatment planning, and guidance expands [1]. This rise in demand is reflected in growing numbers of examinations, while overall waiting times for imaging services are increasing [2]. Including MRI in screening pathways would dramatically increase the demand for MRI capacity [3,4,5]. In the Goteborg-2 trial, second-stage MRI testing following PSA pre-selection included 6.7% and 17.9% of all invited men in the first round when PSA thresholds of 3.0 and 1.8 ng/mL were applied [3,6].

Multiple strategies can reduce the demand for MRI in diagnosis and screening, including using biomarkers [7]. Efficiency can be increased by acquiring only the necessary MRI sequences for the diagnostic task [8], such as omitting dynamic contrast-enhanced imaging [9,10]. Additional measures include accelerated MRI acquisition techniques, where time savings can lead to higher patient throughput [11]. Regardless of the strategies used, it is important to maintain or improve image quality and reduce the need for repeated acquisitions and patient recalls.

There are two main conventional accelerated MRI techniques: Parallel Imaging (PI) and Compressed Sensing (CS). PI uses the sensitivity profile of coil arrays to speed up image acquisition, while CS utilizes incoherent sampling of sparse information in MR images. Both techniques are commonly used in clinical applications, particularly T₂w imaging [12]. Nevertheless, the signal-to-noise ratio and appearance of artifacts restrict the extent of acceleration achievable with these methods.

Deep learning reconstruction (DLR) is a technique that accelerates the process of MRI scanning by saving time, improving quality, and correcting motion [13]. DLR is offered by all major vendors of MR scanners [14] and can be used alongside conventional accelerated MRI methods. A notable aspect of DLR is its computational efficiency [15]. There are different approaches to implementing DLR in the reconstruction process. Image-based learning can improve the quality of reconstructed images after using conventional Inverse Fourier Transform (IFT) (Figure 1). K-space learning can enhance partially sampled raw data before reconstructions with IFT. Direct mapping achieves improvement and reconstruction in one step, using raw data as input to produce an improved image without IFT.

This review examines deep learning (DL)-based reconstruction approaches aimed at either accelerating the acquisition or enhancing the quality of prostate MRI. First, sequence-specific acceleration techniques are evaluated. Second, accelerated screening protocols are reviewed. Third, image quality metrics are analyzed. Finally, we discuss the opportunities and challenges associated with future trials utilizing DL acceleration for protocol personalization in MRI.

2. Materials and Methods

This systematic review was registered in the PROSPERO International Prospective Registry of Systematic Reviews (CRD420251032136), where the protocol is available, and follows the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-analysis for Diagnostic Test Accuracy) guidelines (Appendix A) [16,17]. The electronic databases PubMed, Web of Science, and Google Scholar were searched for primary publications in English assessing DLR for MRI of the prostate published up to 17 March 2025. The first 400 entries from Google Scholar were included. The search terms were: prostate AND MRI AND (deep learning OR DL OR artificial intelligence OR AI) AND (reconstruct* OR sequence).

Two researchers (SU and GA) screened titles, abstracts, and full texts for eligibility. References of included reports were searched for dataset completion. Discrepancies were resolved by consensus. Original reports using artificial intelligence (AI) to improve prostate MRI image quality and/or accelerate acquisition in humans were eligible for inclusion. Conference abstracts, case reports, and non-English reports were excluded. Contact with the authors was sought if reports were not available.

Information from the individual reports was transferred into a data collection form by one reviewer and verified by the second reviewer (Appendix B). Data were included as presented in the included studies. First, sequence-specific acceleration techniques were assessed. Second, accelerated screening protocols were reviewed. Third, image quality metrics were analyzed. Narrative synthesis was used to summarize the effects of DLR on image quality. The reduction in acquisition time achieved with DLR was summarized with descriptive statistics. Narrative synthesis summarized the diagnostic performance of human readers and AI with DLR. One reviewer performed and one verified risk of bias assessment using QUADAS-2.

Quality metrics comparing reference and test sequences were recorded in the data extraction instrument. Due to the high methodological heterogeneity of the studies and the use of numerous quality metrics (Appendix B), quantitative synthesis was restricted to comparing the subjective overall image quality among the studies. The subjective image quality ratings for every reconstruction were recorded (e.g., median and inter-quartile range, mean and standard deviation, or frequency tables). Additionally, the relative acquisition time of the DLR sequence relative to the conventional comparator was recorded.

Consistent with most of the included studies, image quality was treated as a pseudo-continuous variable. The within-group standardized mean difference served as the effect size, considering the varying number of levels in the Likert scales for quality assessment and the paired nature of observations. The correlation between the paired observations, necessary for estimating the standard error, was determined empirically as r = 0.85 on in-house data.

The effect size could not be calculated directly for studies reporting median and interquartile range. To increase the number of studies in a secondary meta-analysis, the means and standard deviations of image quality were computed for these studies as described by Wan et al. This method assumes a normal distribution of the measurements [18]. The inclusion of these additional studies was assessed using a sensitivity analysis. An additional sensitivity analysis was conducted, excluding studies at a high risk of bias.

A three-level, random-effects meta-analysis computed the summary effect size, taking multiple, non-independent observations on a single cohort into account. The meta-analysis was performed separately for T₂w imaging and diffusion-weighted imaging (DWI) and implemented with the “metafor” package (Version 4.8-0) in R language for statistical computing (Version 4.4.1) [19]. A funnel plot visualized the risk of publication bias. The I² statistic estimated the heterogeneity among the studies that was not attributable to random sampling. Appendix E includes a more detailed description of the statistical methods.

3. Results

The search identified 510 unique records, 478 of which were excluded. In particular, 309 investigated unsuitable outcomes, i.e., they did not assess the impact of deep learning image reconstruction on image quality or acquisition time. Examples include studies focusing on using AI for prostate cancer detection with MRI, image segmentation, radiomics analyses, or genomic profiling. Fifty-five studies were classified as evaluating the wrong population because they investigated conditions other than prostate cancer. Examples include imaging studies related to orthopedic, cerebral, or non-prostate pelvic diseases. Eventually, 33 studies (Figure 2) fulfilled the inclusion criteria, investigating possible image quality improvements for individual sequences [20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52], the application of DLR in abbreviated protocols [30,37,40], and the utility of DLR for radiotherapy [52]. Twenty-five reports investigated DLR for T₂w imaging, while 11 assessed DLR for DWI (Figure 3). Studies reported on men undergoing MRI for PCa diagnosis (18), patients with biopsy-proven PCa (3), patients undergoing MRI for multiple indications (11), or healthy volunteers (1). One study [42] included publicly available data from the fastMRI initiative, a dataset of T₂w and DWI containing the k-space data of 312 participants [53,54], complemented by data obtained from the vendor. The remaining 32 studies included only in-house data, with nine stating that the data were available from the authors upon request and one indicating that the data were unavailable.

Studies on T₂w imaging and DWI investigated image-based learning (12), k-space learning (6), or direct mapping (14). Algorithms were developed in-house (9) and by vendors (23), 61% of which were product sequences. Studies most commonly employed 3T machines (85%) from a variety of vendors: Siemens (15), GE (7), Philips (5), Canon (3), United Imaging (1), and Elekta (1) (Figure 3).

3.1. Risk of Bias Assessment

Overall, the risk of bias was low, although seven studies were identified as having a high risk of bias or reduced applicability related to participant selection. This was due to the inclusion of participants with lesions visible on conventional MRI [28,39], those who had undergone biopsy or subsequently underwent prostatectomy [35,36,41,48], or healthy volunteers [44]. These cohorts were either enriched in patients with csPCa or consisted of younger participants with less csPCa. Therefore, they only partially represent clinical target populations. One study was at a high risk of bias in performing the reference test (conventional sequences) and index test (DLR sequences) [34]. The unblinded side-by-side comparison of the two sequences may result in expectation bias, where the readers favor one of the techniques, and anchoring bias, where previous observations influence subsequent evaluations. Appendix D includes the full risk of bias assessment.

3.2. Deep Learning Reconstruction

3.2.1. DLR in T2w Imaging

DLR was employed to accelerate T₂w imaging or improve the image quality sequences without additional acceleration. Nine studies applied DLR without additional acceleration compared to the reference T₂w sequence. DLR achieved a higher image quality in seven reports, while two observed a lower quality. Seven studies used objective image quality metrics. The most commonly calculated image quality metrics were the SNR and the contrast-to-noise ratios (CNR) in three and four studies, respectively. One study used other metrics. The objective image quality improved in five studies, was unaltered in one study, and deteriorated in another study when acquisition times were constant.

DLR can potentially accelerate image acquisition while maintaining or increasing image quality. This was investigated with accelerated acquisitions (15) or simulations (3;

Table 1. Summary of acceleration approach (simulated or acquired) and domain of the input data (image space, k-space, or k-space with direct mapping).

	T2w			DWI
	Image	k-Space	Direct Mapping	Image	k-Space	Direct Mapping
Simulated	2	0	1	2	1	2
Acquired	3	2	10	0	0	2
Not applicable	4	3	0	2	1	1

). The acquisition time was reduced to a minimum of 13–64% of the reference sequence, representing a 1.56–7.7-fold acceleration. Acceleration did not lead to a deterioration in objective image quality metrics in any of the studies. Subjective preference at low acceleration factors favored DLR. At 3-fold and higher acceleration, DLR was of comparable quality to conventional sequences.

Six studies evaluated the effect of DLR on clinically relevant tasks. These included cancer detection by AI algorithms and human diagnostic performance, and detecting extraprostatic extension (EPE). AI algorithms trained on conventional MRI data performed worse on DL-reconstructed T₂w data [47,50]. DLR had a less significant impact on human diagnosis, showing comparable performance for transition zone lesions, despite a 3-fold acceleration [35]. Another study with 10 patients showed similar diagnostic performance [30]. Evidence on EPE detection varied: one report showed similar performance despite a 37–70% reduction in acquisition time [35,41], while another showed decreased performance with high-resolution DLR T₂w imaging [36].

Evaluation cohorts included a median of 46 participants (IQR 30–96). The study by van Lohuizen et al., a multicenter trial of three centers in the Netherlands, employed the largest testing cohort, with 306 participants [50]. It used reconstructed images and Fast Fourier Transform (FFT) to obtain sub-sampled k-space-like data that were sub-sampled, avoiding the prospective collection of k-space data. Some ADC values derived from DLR images varied from the reference but showed a strong linear correlation. Yet, these differences were typically modest and within the test–retest limits defined by the Quantitative Imaging Biomarkers Alliance (QIBA) profile 2024 [55].

3.2.2. DLR in Diffusion-Weighted Imaging

Without additional acceleration, the subjective image quality of DLR DWI was comparable in three studies and improved in two studies. All five studies found a higher signal-to-noise ratio (SNR) in DLR images (1.24–2.17-fold). The studies had no other shared objective image quality metric.

DLR was employed to accelerate image acquisition in eight studies. Acquisition time reductions ranged between 13% and 69% of the reference sequences. The subjective image quality was independent of the degree of acceleration. However, retrospective sub-sampling of k-space simulated acceleration in five studies, and only three studies acquired accelerated sequences, potentially reducing gains in image quality from reduced motion artifacts. The effects of DLR in DWI on downstream tasks remain unexplored. Models were evaluated on relatively small cohorts with a median of 51 participants (IQR 36–60), lowering clinical generalizability.

3.2.3. DLR in AI-Accelerated Screening/Diagnostic Protocols

Three studies investigated complete accelerated bi-parametric protocols [30,37,40], comparable to proposed screening protocols.

In a retrospective study, Johnson et al. simulated undersampling of axial and coronal T₂w and axial DWI sequences, validating a simulated acceleration from 11:48 min to 3:12 min (− 73%) at maintained image quality in 20 participants. No difference in diagnostic performance was detected, albeit only a few lesions were present in the cohort [30].

Lee et al. conducted a retrospective analysis of 40 prospectively recruited patients who underwent axial T₂w and DWI. Standard of care imaging, T₂w with a reduced number of averages and standard or higher resolution, and DWI with a reduced number of averages were reconstructed with different DLR strengths. With all of the sequences, DLR increased SNR and CNR ratios. Two radiologists rated the overall image quality as similar to the standard of care for most reconstructions. Only the accelerated T₂w sequence reconstructed with medium-strength DLR showed a higher quality [37].

Oerther et al. studied 77 men suspected of prostate cancer using an accelerated and abbreviated bi-parametric MRI protocol with T₂w imaging and DWI (3:28 min) compared to standard mpMRI with 3-planar T₂w TSE and DCE (25:45 min). The abbreviated protocol had better image quality, higher inter-rater agreement in PI-RADS 2.1 scoring, and comparable diagnostic performance to the standard protocol [40].

3.2.4. DLR in PI-RADS Assessment

Thirteen studies assessed the effect of DLR on PI-RADS scoring, including studies investigating bi-parametric screening protocols (DLR T₂w and DLR DWI; 3), DLR T₂w with conventional DWI (9), or DLR DWI with conventional T₂w (1). Most studies found good-to-excellent agreement or no significant difference between PI-RADS scoring performed on conventional and DLR images. Only one study observed significant upgrading with DLR DWI [29]. Inter-reader agreement increased with DLR in three studies [25,29,50]. In a prospective study of DLR for T₂w imaging, Bischoff et al. observed a change in PI-RADS scores in 6% of patients that reduced the number of PI-RADS 3 cases by 78–86% through up- or down-scoring [21].

In a study comparing conventional multi-parametric MRI to DLR bi-parametric MRI, one of two readers had a lower specificity for lesions ≥ PI-RADS 3 and a lower sensitivity for lesions ≥ PI-RADS 4 on DLR bi-parametric MRI [43]. Only one study offered sufficient details to calculate an effect size [21]. Therefore, a meta-analysis of diagnostic performance was not possible.

3.3. Methods for Image Quality Assessment

3.3.1. Objective Image Quality Metrics

Twenty-six studies employed objective image quality metrics. However, the specific metrics used in the studies varied widely (

Table 2. Summary of objective and subjective image quality metrics used in the included studies.

Objective Image Quality Metrics	Subjective Image Quality Metrics
- Signal-to-noise ratio (15)	- Overall image quality
- Contrast-to-noise ratio (13)	- Noise
- Edge rise distance / Slope profile (6)	- Artifacts
- Structural similarity index (6)	- Sharpness
- Peak signal-to-noise ratio (5)	- Conspicuity
- Root mean squared error (4)	- Diagnostic confidence
- Performance of diagnostic AI (3)
- Feature similarity index (1)
- Perceptual index (1)
- Variance (1)
- Gaussian log likelihood (1)
- Edge keeping index (1)

). Image quality metrics most commonly included SNR, CNR, and edge rise distance/slope profile in 13, 12, and 6 studies, respectively. The most commonly used fully referenced quality metrics were structural similarity index, peak SNR, and root mean square error. Those metrics compare DLR images against a high-quality ground truth.

3.3.2. Subjective Image Quality Metrics

Twenty-nine studies assessed subjective image quality with one to five readers (median 2). Quality metrics, while differently defined, fell into one of six categories: overall image quality, noise, artifacts, sharpness, conspicuity of organs, and diagnostic confidence. Binary classifications, 4- and 5-tier Likert classifications were used.

Appendix B lists all subjective and objective image quality metrics for individual studies.

3.4. Quantitative Synthesis

Twenty-six studies reported subjective overall image quality and were potentially eligible for inclusion in the meta-analysis. Five studies did not report sufficient details to obtain an effect size. Two studies were excluded because they reported an interquartile range of 0, which did not allow for approximating a standard error. The final meta-analysis included eight studies in the primary analysis. An additional 11 studies with effect sizes computed indirectly were included in the extended cohort for the sensitivity analysis (Appendix F). The studies reported 26 comparisons in the primary and 52 in the extended analysis. One study reported a combined quality score for T₂w imaging and DWI included in the meta-analysis of both sequences.

The meta-analysis revealed a non-significant trend towards a higher image quality with DLR on T₂w imaging (summary effect size of 0.49, 95% confidence interval −0.39–1.39, p = 0.25, df = 13, Figure 4a). However, the overall heterogeneity of between-study comparisons was high, with an I² of 77%. The heterogeneity of within-study comparisons was moderate at I² = 22%, resulting in a total heterogeneity of I² = 98%. The funnel plot reveals some asymmetry favoring DLR, indicative of possible publication bias (Figure 5). The sensitivity analysis, including studies where the effect size was approximated from median and interquartile range, showed a borderline significant trend towards higher image quality for DLR (summary effect size of 1.38, 95% confidence interval 0.00–2.04, p = 0.05, df = 32). However, this increase in quality was determined by two studies with extreme results (Appendix F). Image quality was not associated with the relative acquisition time (Figure 6a).

The analysis of DWI revealed no significant association of image quality with the reconstruction method (summary effect size of 0.67, 95% confidence interval −0.41–1.74, p = 0.21, df = 13, Figure 4b). The overall heterogeneity of between-study comparisons was high, with an I² of 78%. The heterogeneity of within-study comparisons was moderate at I² = 20%, resulting in a total heterogeneity of I² = 99%. The funnel plot showed a moderate asymmetry favoring DLR (Figure 5), indicative of possible publication bias. The sensitivity analysis, including studies where the effect size was approximated from median and interquartile range, did not change the results (summary effect size of 0.62, 95% confidence interval −0.26–1.50, p = 0.16, df = 19, Appendix F). Meta-regression showed no significant association between the relative acquisition time and image quality Figure 6.

4. Discussion

This systematic review evaluated the rapidly increasing literature on DLR for prostate MRI. Most of the research took place in the screening and early diagnostic setting. However, cohorts were frequently heterogeneous. The assessment of DLR primarily focused on image quality evaluation, and research on the effects of DLR on downstream tasks is only just beginning to accrue. Methodological heterogeneity and considerable variability in quality metrics limited the scope of quantitative comparisons. The potential of DLR for prostate imaging has been recognized and reaches beyond diagnostic imaging with initial research on the use of DLR for treatment guidance in radiotherapy.

4.1. Benefits of DLR

DLR can impact MRI protocol development and aid in creating MRI screening strategies for prostate cancer. This review demonstrates that DLR can speed up T₂w and DWI acquisition of the prostate by up to 8-fold without sacrificing image quality [50]. A meta-analysis of the subjective image quality revealed no significant difference between conventional and DLR images among the included studies. However, there was considerable heterogeneity among the studies, with some aiming for image quality improvement and some for accelerated scanning time. Faster scans improve patient comfort and reduce motion artifacts. DLR can enhance image quality on lower-specification MRI systems without longer acquisition times. Cost-effectiveness studies indicate that DLR increases scanner capacity more effectively than buying another MRI system or extending weekend hours [11].

4.2. Risks of DLR

While image quality in conventional reconstruction was a trade-off between acquisition time, SNR, and resolution, additional factors play into DLR (Figure 7). One key risk is hallucinations, the addition or removal of realistic-looking image details, which may arise when there are multiple plausible solutions to the inverse image reconstruction problem. A deep learning model, influenced by the priors learned from the training data, may favor solutions introducing hallucinated features. The risk of hallucinations increases with stronger acceleration, i.e., decreasing sampling rates [56]. Reduced image noise may also reveal artifacts that are not ordinarily visible, leading to false-positive findings. A second phenomenon is instability, where small perturbations in the input (e.g., noise) lead to severe artifacts or degradation of image quality [57]. Importantly, hallucination and instability are inevitable in many cases of DLR. Achieving perfect accuracy, stability, and hallucination-free reconstructions simultaneously is not possible [58]. Nonetheless, DLR and conventional acceleration techniques require similar caution in their clinical implementation.

Further artifacts specific to DLR include over-smoothing, which can mimic the charcoal sign in the transition zone, leading to false positives. Conversely, the loss of subtle findings adjacent to the capsule may cause false negatives. These artifacts depend on algorithm settings, vendors, scanners, and patient anatomy.

4.3. Effect of DLR on PI-RADS Assessment and Downstream Clinical Tasks

Image quality is only a surrogate marker for the aim of imaging: improving outcomes in a diagnostic or therapeutic pathway. Therefore, it is essential to evaluate the effect of DLR on downstream tasks such as the diagnosis of prostate cancer, AI-assisted lesion detection, biopsy guidance, or automated segmentation for radiotherapy (Figure 1). Initiatives to probe the effect of DLR on downstream tasks were ongoing.

There was no evidence that DLR would significantly impact PI-RADS scoring for human readers. The agreement between conventional and DLR images is generally high, and the significant difference in one study was modest in magnitude [40]. As only one comparison used DLR in T₂w imaging and DWI simultaneously, data on fully DLR protocols are still scarce. Nonetheless, no evidence suggested that DLR significantly impacts PI-RADS scoring.

It is also relevant to go beyond comparing two imaging outcomes (e.g., PI-RADS). Instead, we should relate outcomes from imaging to a reference standard, e.g., follow-up or histological diagnosis of PCa. Among the studies that evaluated DLR on a downstream task not based on an imaging outcome (e.g., detection of histologically proven cancer or extraprostatic extension), the performance of conventional and DLR images was similar [29,30,34,35,41,48]. Where significant differences occurred, they were generally small [36]. This non-inferior evidence must still be considered preliminary, as the studies were retrospective, often included small patient cohorts, were at a high or uncertain risk of selection bias, and used biopsies guided by standard MRI as the reference standard. Biopsy guidance with DLR sequences remains unexplored.

ADC values derived from DLR may exhibit slight deviations from those obtained using conventional imaging techniques. These differences can be attributed to several factors: (1) denoising effects, where noise reduction—particularly at high b-values—can unmask underlying signal; (2) the characteristics of the DLR algorithm and its training data; (3) variations in sequence design and imaging protocols; and (4) tissue-specific effects. While the clinical significance of these differences is likely limited, given that relative ADC contrasts between tissues are generally preserved and the absolute changes are minor, sequence-specific validation is advisable before quantitative analysis.

4.4. Effect of DLR on Diagnostic AI

Diagnostic AI for PCa detection on MRI has been a focus of intense research, and models for clinical use are available from multiple vendors. Therefore, the combined diagnostic performance of DLR images as input for diagnostic AI algorithms is of great importance.

Van Lohuizen et al. and Tong et al. investigated the effect of DLR of T₂w images on diagnostic prostate AI [47,50]. While on objective and subjective metrics, image quality was higher than or comparable to the reference standard in both studies, the performance of diagnostic AI on DLR images was inferior. This highlights the challenge posed by the out-of-domain nature of DLR images. One solution is to incorporate DLR images into the training data for AI. Nonetheless, use cases of AI where multiple AI algorithms interact may lead to unexpected outcomes (the double black box phenomenon), warranting careful evaluation in clinical practice.

4.5. Heterogeneity in the Evaluation of DLR Image Quality

Various metrics were employed for subjective and objective image quality assessment, limiting the comparability of the results across the studies included in this systematic review. An objective quantification of image quality is attractive as it may allow a comparison of outcomes across different DLR algorithms. However, the apparent precision of a numerical metric may correlate only poorly with the visual quality of an image. The SNR and CNR were the most common quality metrics. While both relate to image noise and signal, they neglect sharpness and image distortions. Similarly, the most common fully referenced metrics, peak SNR and structural similarity index measure (SSIM), agree poorly with radiologists’ subjective image quality assessments [59]. Furthermore, they are strongly dependent on image normalization [60].

No single image quality metric is sufficiently sensitive to all possible causes of image distortions. In response to this, Dohmen et al. suggested a combination of SSIM, learned perceptual image patch similarity (LPIPS), mean squared error (MSE), and normalized mutual information (NMI) to cover a large range of possible distortions [60]. Still, using fully referenced metrics will be unable to detect improvements in the quality of DLR images compared to the reference standard.

4.6. DLR Beyond Diagnosis

The increasing demand for prostate MRIs will likely require shorter and more personalized protocols that will have to be adapted more closely to downstream tasks. Consequently, novel sequences need to be evaluated in concert with the intended use case. For example, a diagnostic protocol requires high sensitivity to identify clinically significant cancer. Meanwhile, a simulation/planning scan for radiation oncology might weigh accurate anatomical representation higher, as this will impact treatment planning and dose delivery to known lesions. Finally, MRI for adaptive radiotherapy serves as the confirmation of existing treatment plans. Therefore, radiation oncologists may accept a slightly reduced image quality if the shorter acquisition time reduces patient motion and discomfort.

4.7. Limitations

This systematic review has some limitations. The heterogeneity among the included studies restricted the scope of the quantitative synthesis to the subjective overall image quality. This analysis confirmed the high methodological heterogeneity among the studies. Studies varied in terms of the included patient population, the operating point of the AI models (quality improvement vs. acceleration), and metrics for image quality assessment. Additional variation was introduced by different scanners from different vendors, with variable field strengths, and different clinical outcomes. Such variation represents the heterogeneity among medical centers, as the centers’ role in the healthcare system, long-term infrastructure investment, and software/hardware integration differ. It is also representative of a multi-faceted challenge where image quality is only the first layer. Only a few studies investigated the impact of DLR on the diagnostic performance of prostate MRI or biopsy guidance as clinically relevant endpoints. Finally, the review only included studies published in English. This decision was based on the limited proficiency of the research team in other relevant languages, which would have hindered devising a search strategy and assessing manuscripts consistently and reliably across languages.

4.8. Summary and Recommendations

Box 1 summarizes the principal findings of this systematic review. Box 2 presents recommendations and suggestions for the implementation of DLR for prostate MRI based on the findings of this systematic review.

Box 1. Principal findings of the systematic review.

Image quality in T₂w and DWI sequences

• Non-accelerated DLR improves the subjective and objective image quality of T₂w sequences, with maintained quality at 3-fold and stronger acceleration.
• DLR achieves a higher SNR in DWI sequences.
• Diagnostic AI performs invariably less well on DLR images, even if objective and subjective image quality metrics were higher.

DLR screening protocols and diagnostic performance

• DLR may reduce acquisition times without a loss in image quality.
• PI-RADS scoring is comparable between conventional and DLR sequences.
• Bi-parametric DLR protocols show similar diagnostic accuracy in small study cohorts.

Methods image quality assessment

• Common image quality metrics include no-reference and fully-referenced metrics (structural similarity index, peak SNR, and root mean square error).
• Subjective image quality metrics assess overall image quality, noise, artifacts, sharpness, conspicuity of structures, and diagnostic confidence.

Risks and limitations of DLR

• Stronger acceleration increases the risk of hallucinations and instability.
• Perfect accuracy, stability, and hallucination-free reconstructions are not achievable with DLR, requiring caution in clinical implementation.

Box 2. Recommendations and suggestions on the implementation of DLR.

Recommendations:

• Select the acceleration factor of DLR judiciously to balance the desired acquisition time savings with the risk of hallucinations and instability [56].
• Implement deep learning image reconstruction in coordination with the entire clinical team, including referring physicians if necessary, e.g., by means of side-by-side review sessions.
• Monitor the compatibility of DLR with downstream workflows such as diagnostic AI [47,50].
• Adapt DLR to the specific clinical context. The trade-off between image quality and acquisition time may vary between use cases [52].
• Monitor diagnostic performance and biopsy yield, especially when biopsy decision making transitions from conventional to DLR sequences [21].

Considerations:

• Implement DLR in a step-wise fashion rather than full DLR protocols.
• Implement DLR in addition to conventional sequences initially and while fine-tuning sequence parameters to achieve the desired diagnostic quality.
• Fully replace secondary sequences (coronal and sagittal T₂w, axial T₁w and DCE) before replacing axial T₂w and DWI.
• Standardize parameters for image quality assessment or find a consensus on which one should be used.
• Investigate the effect DLR would have had on biopsy decisions in patients undergoing conventional and DLR imaging.
• Perform a comparative analysis of diagnostic AI using conventional and DLR-acquired images to assess consistency and potential performance shifts.
• Record the use of DLR in reports to facilitate audits and AI monitoring.

5. Conclusions

Image quality analyses indicate that DLR can accelerate high-quality prostate MRI in various use cases. Future research should prioritize prospective, reader/blinded, and paired study designs with histological validation as a reference standard to determine the effect of DLR on diagnostic accuracy. Developing task-specific protocols will be key to successfully implementing DLR, particularly when additional AI models are applied to images further downstream in the analysis pathway.