Next Article in Journal
Strain Sensor-Based Fatigue Prediction for Hydraulic Turbine Governor Servomotor in Complementary Energy Systems
Previous Article in Journal
A ST-ConvLSTM Network for 3D Human Keypoint Localization Using MmWave Radar
Previous Article in Special Issue
Prediction of Germline BRCA Mutations in High-Risk Breast Cancer Patients Using Machine Learning with Multiparametric Breast MRI Features
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Sensors
Volume 25
Issue 18
10.3390/s25185858
sensors-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Cross-Scanner Harmonization of AI/DL Accelerated Quantitative Bi-Parametric Prostate MRI †

by
Dariya Malyarenko
1,*,
Scott D. Swanson
1,
Jacob Richardson
1,
Suzan Lowe
1,
James O’Connor
1,
Yun Jiang
1,
Reve Chahine
1,
Shane A. Wells
1,2,‡ and
Thomas L. Chenevert
1,‡
1
Department of Radiology, University of Michigan Medical School, Ann Arbor, MI 48109, USA
2
Department of Urology, University of Michigan Medical School, Ann Arbor, MI 48109, USA
*
Author to whom correspondence should be addressed.
†
This article is a revised and expanded version of Harmonization of AI/DL accelerated quantitative bi-parametric prostate MRI: demonstration in multiparametric phantom and patients. PP103. In Proceedings of the Annual Meeting of the International Society for Magnetic Resonance in Medicine, Honolulu, HI, USA, 10–15 May 2025.
‡
These authors contributed equally to this work and share co-senior authorship.
Sensors 2025, 25(18), 5858; https://doi.org/10.3390/s25185858
Submission received: 5 August 2025 / Revised: 12 September 2025 / Accepted: 17 September 2025 / Published: 19 September 2025
(This article belongs to the Special Issue Applications of Biomedical Imaging and Sensing Technologies in Disease Diagnosis)
Browse Figures
Versions Notes

Abstract

Clinical application of AI/DL-aided acquisitions for quantitative bi-parametric (q-bp)MRI requires validation and harmonization across vendor platforms. An AI/DL-accelerated q-bpMRI, including 5-echo T2 and 4-b-value apparent diffusion coefficient (ADC) mapping, was implemented on two 3T clinical scanners by two vendors alongside the qualitative standard-of-care (SOC) MRI protocols for six patients with biopsy-confirmed prostate cancer (PCa). AI/DL versus SOC bpMRI image quality was compared for MR-visible PCa lesions on a 4-point Likert-like scale. Quantitative validation and protocol bias assessment were performed using a multiparametric phantom with reference T2 and diffusion kurtosis values mimicking prostate tissue ranges. Six-minute q-bpMRI achieved acceptable diagnostic quality comparable to the SOC. Better SNR was observed for DL/AI versus SOC ADC with method-dependent distortion susceptibility and resolution enhancement. The measured biases were unaffected by AI/DL reconstruction and related to acquisition protocol parameters: constant for spin-echo T2 (−7 ms to +5 ms) and ADC (4b-fit: −0.37 µm2/ms and 2b-fit: −0.19 µm2/ms), while nonlinear for echo-planar T2 (−37 ms to +14 ms). Measured phantom ADC bias dependence on b-value range was consistent with that observed for PCa lesions. Bias correction harmonized lesion T2 and ADC values across different AI/DL-aided q-bpMRI acquisitions. The developed workflow enables harmonization of AI/DL-accelerated quantitative T2 and ADC mapping in multi-vendor clinical settings.

1. Introduction

Current clinical management of patients with prostate cancer (PCa) relies on costly and painful multi-core needle biopsies performed for both diagnosis and active surveillance. Due to the relatively low occurrence of high-risk cancers, 80–90% PCa patients could potentially be spared unnecessary biopsies [1,2]. The standard-of-care (SOC) multiparametric (mp)MRI improves quality of life, [3], and qualitative interpretation according to Prostate Imaging Reporting and Data System (PIRADS) [4] provides excellent sensitivity for PCa detection [5,6]. However, modest specificity of subjective PIRADS scores necessitates biopsy confirmation for PRADS > 2 lesions, including a large portion of indolent cancers. For lesion assessment in the peripheral zone where the majority of PCa occur [2,7], bi-parametric (bp)MRI is typically sufficient. In addition to the T2-weighted anatomical scan, the diffusion weighted imaging (DWI) component of the bpMRI SOC exam includes calculation of an apparent diffusion coefficient (ADC) and qualitative examination of high b-value (>1400 s/mm2) images for signs of impeded diffusion. ADC maps are routinely generated on the scanner with different hospitals using variable b-ranges that affect derived values [5,8], complicating quantitative ADC interpretation.
Imaging studies that add assessment of quantitative diffusion parameter [8,9,10] and T2 tissue relaxation [11,12,13] improve specificity for clinically significant (cs)PCa [8,9,14] that tends to have lower ADC and T2 values, which may reduce the need for biopsies. However, comprehensive quantitative bpMRI studies typically require lengthy (multi-b and multi-TE) custom acquisitions [12,13,14,15] that limit their practical utility for clinical SOC examinations. PCa MRI protocol optimization is subject to a balance between acquisition time, spatial resolution, and signal-to-noise ratio (SNR). Hence, diffusion and relaxation metric values derived from abbreviated acquisition protocols are inevitably biased, depending on the applied models [8,10,11,12,15,16,17] and acquisition settings (e.g., b-value or TE ranges) [14,18,19]. These biases would also vary across MRI vendor platforms used in multi-site clinical trials and for longitudinal follow-up of patients on active surveillance in clinical settings [5,8,14]. Thus, measured ADC and T2 values require harmonization for quantitative comparison across studies and vendor platforms to establish uniform thresholds for q-bpMRI metrics of csPCa [5,6,9] and enable translation to clinical practice [14].
Emerging artificial intelligence (AI) deep learning (DL) reconstruction and denoising methods hold potential to substantially accelerate bpMRI and improve image quality [20,21,22]. However, their clinical utilization for quantitative mapping is hindered by undetermined biases across vendor AI/DL implementations that are trained on proprietary image collections with unknown acquisition parameters [23,24]. AI/DL models are highly susceptible to training set biases and necessitate rigorous optimization and verification [23,24]. In clinical settings, AI/DL-accelerated reconstruction tools are typically built on the scanner consoles as black-box pre-trained models and filters with limited options for adjustment parameters varying across MRI systems [23]. Furthermore, clinical scan acquisition parameters are often modified in real time on a per-patient basis, which is incompatible with prospective protocol standardization [14].
Quantitative validation of the front-end AI/DL-aided reconstruction and image processing provided by scanner vendors is thus particularly challenging and requires a reference platform compatible with the clinical workflow. According to QIBA and FDA guidelines for quantitative imaging devices [25,26,27], physical phantoms with known true parameter values supply the most practical means for assessment of mpMRI protocol accuracy and multi-system harmonization by correction of technical biases. Given the multitude of options and combinations of MRI acquisition and processing parameters, a comprehensive verification of AI/DL-aided quantitative relaxation and diffusion mapping protocols in the clinical environment requires an independent mpMRI reference standard of realistic complexity.
We have recently developed a quantitative multiparametric phantom for objective bias assessment of relaxation and diffusion parameters in prostate tissue relevant ranges [28]. This realistic image reference system is fully compatible with clinical scan protocols and allows objective bias assessment and correction independent of AI/DL protocol parameters and vendor algorithms. This work demonstrates the use of the mp-phantom for quantitative evaluation of biases from AI/DL-aided prostate q-bpMRI acquisition protocols and fit models. The purpose of this study was to assess acquisition protocol biases versus the added contribution of AI/DL reconstruction methods to enable quantitative T2 and ADC harmonization across multiple vendor platforms.

2. Materials and Methods

2.1. AI/DL-Accelerated Quantitative T2 and ADC Mapping Protocols

The vendor-provided prototype artificial-intelligence (AI) deep-learning (DL) reconstruction and denoising methods [20,21] were implemented for prostate apparent diffusion coefficient (ADC) and transverse relaxation time (T2) mapping on 3T Vida (Sys1: Siemens, Erlangen, Germany) and 3T Ingenia (Sys2: Philips, Best, The Netherlands) clinical scanners. According to vendor-provided descriptions, both methods have used the image-trained AI/DL network models for undersampled compressed sensing reconstruction. The total acceleration was a factor of six with system-default denoising, super-resolution (sr), or partial-Fourier (pf) settings. For multi-echo spin-echo (MESE) [29] or echo-planer-imaging (MEEPI) [12], T2 mapping used five echo-times (TE) over the range recommended in the literature (Table 1). For ADC mapping, diffusion weighted imaging (DWI), four b-values (with-averaging) were acquired as in institutional SOC practice (Table 1).

2.2. Patient Studies

The prospective, single-center study was IRB-approved and HIPAA-compliant. Informed consent was obtained from all participants. The AI/DL-accelerated protocols were added to SOC exams for six patients (enrolled June–December 2024) on active surveillance for PCa with lesion pathologic grade initially obtained by needle biopsy [3]. Four patients had MR visible lesions (two PIRADS 4 and two PIRADS 5) with biopsy Gleason 3 + 3. The clinical SOC1/SOC2 scans (45 min) included about 15 min of bi-parametric (T2-weighted and 4b-value DWI) acquisitions (Table 1) [30]. T2-mapping was not a part of SOC and was only included for AI/DL-aided acquisitions. The reconstructed voxel sizes for AI/DL images were matched to interpolated SOC values. Two raters (a clinical radiologist with 18 years and a radiology fellow with 7 years of experience) independently qualitatively compared [31,32] the SOC T2w, high-b DWI and ADC images to the corresponding AI/DL outputs for diagnostic quality (capsule demarcation and zonal anatomy contrast), resolution, signal-to-noise-ratio (SNR) and distortion using four-point Likert-like scale (as 4: better, 3: similar, 2: worse/acceptable or 1: worse/nondiagnostic). The EPI distortion was assessed visually with respect to the SOC T2w images. Lesion SNR was evaluated by visual inspection with respect to background image noise, and resolution was assessed from visual conspicuity of prostate boundaries. The overall performance for individual contrasts was summarized by median scores, and agreement between raters was measured by correlation coefficient (R) and individual score difference range.

2.3. Phantom Measurements

A previously developed multiparametric phantom consisting of four physical layers of hydrogels providing tissue-mimicking ADC, kurtosis (K), and T2 values (Supplementary Figure S1B) was used as a reference with relaxation and diffusion parameters pre-calibrated from 8-TE TSE-MESE and 8-b EPI-DWI in a high-SNR head coil as described [28]. After thermalizing overnight, the phantom was scanned at ambient temperatures (Ts = 21.0–23.5 °C) for each patient examination with patient-specific acquisition settings. Phantom temperature (±0.5 °C) was measured using NIST ADC calibration for water and 20% polyvinyl pyrrolidone (PVP20) internal standards (Supplementary Table S1) included in the jar with the multiparametric tissue mimics [28]. T2 and ADC values derived from AI/DL-accelerated protocols were compared to the reference values (Supplementary Figures S1 and S2) for bias estimation. Observed minor ambient temperature variation for phantom scans ensured a limited contribution to measured protocol biases. The diffusion kurtosis (DK) model [16] was used for the diffusion ADC reference standard, and mono-exponential T2 relaxation [11] for T2 reference values. These models have appropriately represented in vivo T2w and DWI signal dependence on TE and b-value ([28], Supplementary Figure S1). A single calibration reference was used for analysis of all clinical protocols. The stability of ADC(DK) reference values [28] was checked with a repeated calibration scan (Supplementary Table S2) and appeared to be within the measurement uncertainty limits (±0.05 µm2/ms).

2.4. Quantitative Bi-Parametric (bp) MRI Analysis

ADC(4b) and MESE-T2 maps were reconstructed on the scanners, while MEEPI-T2 ADC(2b) (b = 100 s/mm2 and b = 800 s/mm2), and ADC(DK) were derived off-line using a mono-exponential model fit of voxel log-signal intensity as a function of TE and b-value, respectively. To mimic clinical workflow, no off-line image normalization, noise reduction, or artifact removal was applied.
The T2 MEEPI mapping [12] was performed according to the mono-exponential model:
l o g ( S T E S 0 ) = − T E T 2
as a function of echo-time (TE). The ADC(2b) fit model was:
l o g ( S b 800 / S b 100 ) = − Δ b × A D C
where Sb800 and Sb100 are the signal intensities with diffusion weighting factor, b = 800 and 100 (s/mm2). The reference diffusion kurtosis model for ADC(DK) was:
l o g ( S b / S 0 ) = − b × A D C + K × ( b × A D C ) 2 / 6
where K is diffusion kurtosis, Sb and S0 are the signal intensities with and without diffusion weighting. All data analysis utilized MATLAB R2019b (Mathworks, Natick, MA, USA).

2.5. Phantom Metrics and Protocol Bias Measurement

Phantom T2 and ADC parameter histograms (Supplementary Figure S2) were generated by manually placing 50 × 15 mm2 rectangular regions-of-interest (ROIs) on the middle slice of the parametric maps (away from vial edges). Bin sizes for phantom T2 and ADC histogram were set to 3 ms and 0.03 µm2/ms, respectively. Measurement errors for histogram peak mean and half-width (HW) were half of the bin sizes. The protocol SNR was assessed from the ratio of the phantom histogram peak maximum to its HW (precision/uncertainty). Bias was quantified as the deviation of the measured mean parameter values from phantom reference values [26,27]:
Bias (T2 or ADC) = (T2 or ADC) − reference (T2 or ADC)
The phantom bias dependence on measured parameter value for patient AI/DL-aided quantitative T2 and ADC mapping protocols on the same system (termed “repeats”) was fit to a common regression model (constant for ADC and quadratic for T2). The bias correction was performed by subtracting the fit bias from the measured metric value for tissue mimics. The agreement among protocols was assessed by standard deviation (SD), and accuracy was assessed by average deviation from reference values before and after correction. The level of significance (p < 0.05) of correction impact on T2 and ADC metrics in phantoms was tested by one-way ANOVA comparison of mean values for tissue mimics before and after bias correction.

2.6. Quantitative Lesion Metrics Harmonization

Lesion ROIs were manually traced on a single slice of each patient’s ADC(4b) maps to avoid artifacts and partial volume. The same ROI was applied for all ADC and T2 measurements of a patient. Bin sizes for patient lesion T2 and ADC histograms were set to 5 ms and 0.05 µm2/ms, respectively. The lesion histogram half-width had a likely contribution from biological heterogeneity, precluding its use for precision and SNR assessment. Mean lesion T2 and ADC parameter values were harmonized across platforms/protocols by bias correction. The lesion bias for mean T2 and ADC values was corrected by subtracting the bias measured for the phantom (Equation (4)) with the patient-specific protocol. The T2 and ADC parameter bias was determined from the corresponding phantom fit bias value for the measured mean lesion value. The phantom-derived bias corrections were compared for measured lesion ADC dependence on b-range. The correction (harmonization) efficacy was assessed by average residual differences across lesions for corrected ADC(2b) and ADC(4b) versus measured ADC(DK).

3. Results

3.1. Qualitative Assessment

The AI/DL accelerated four-b DWI and five-TE T2 scans (Table 1) took less than 6 min on each MRI platform (less than half of the corresponding standard-of-care bpMRI examination time). Among SOC and AI/DL scan protocols, b-value sets were consistent while other acquisition parameters varied (Table 1). Acquired AI/DL T2w in-plane resolution (2 × 2 mm2) and high-b-averages (8) for DWI were lower than SOC1 (0.5 × 0.6 mm2 and 20) or SOC2 (0.4 × 0.7 mm2 and 16). TR parameter varied most between patient scan protocols on Sys1, and for AI-accelerated T2 and ADC on Sys2. Different through-slice resolutions (4 mm and 3 mm) were tested for AI/DL-aided EPI DWI. DL reconstruction protocol on Sys1 was compatible only with EPI, hence MEEPI was implemented for quantitative T2-mapping for patient scans on this system. All SOC scans had adequate diagnostic quality, resolution, and SNR, and an acceptable level of distortion. The same (factor of 6) acceleration and number of b-averages without DL/AI-aided denoising would yield an unacceptable level of noise and distortion in SOC DWI scans.
Four MRI visible lesions were subjectively evaluated by two raters on a 4-pt Likert-like scale for AI/DL-aided reconstruction protocols in comparison to SOC for diagnostic quality, contrast, SNR, resolution, and distortion (Table 2). The overall interrater agreement was good (R = 0.78) with a maximum individual score deviation of ±1. In general, both raters confirmed acceptable diagnostic quality for AI/DL abbreviated protocols, similar to SOC (median Likert-like scores 2–3 for Sys1 and 3–4 for Sys2 patient scans), although individual sequence performance varied. Overall, lesion DWI and T2w contrast features of adequate (acceptable or similar) diagnostic quality were perceived on DL/AI versus SOC bpMRI (Figure 1, Table 2), except for T2w (DLpf-EPI) of Sys1-pt2 with high EPI-distortions. Median performance scores were higher for Sys2 patients and for T2w images derived from AIsr-MESE versus DLpf-MEEPI.
For DWIb1600 and ADC, both DLpf (Sys1) and AIsr (Sys2) methods showed similar or improved SNR compared to SOC for 4 mm slices (scores of 3 and 4). SNR was lower, and acceptable for 3 mm slices, but the resolution of AIsr (Sys2) reconstruction was always better than SOC. Interestingly, the DLpf (Sys1) protocol appeared to be more susceptible to EPI-distortions compared to AIsr (Sys2) reconstruction, in the presence of rectal air (Figure 2). DLpf (Sys1) MEEPI protocol showed generally lower performance for T2w images compared to SOC in resolution, SNR, and distortion. AIsr (Sys2) method was compatible with MESE protocol, leading to similar or better T2w SNR, resolution, and distortion versus SOC images, and would be preferred over T2w MEEPI (Table 2, Figure 2).

3.2. Quantitative Assessment

Figure 3 illustrates an objective quantitative assessment of scan-protocol bias for AI/DL-accelerated acquisitions (used for four MR-visible lesions) with respect to the DK phantom reference (Supplementary Figure S2). The bias measured from the shift in DL/AI-aided protocol histogram peaks from the reference peaks (Supplementary Figure S2, Equation (4)) for tissue mimics is summarized in Table 3. The parameter measurements for additional protocols before and after bias correction are detailed in Supplementary Tables S3 and S4.
Over the reference T2 range (70–170 ms) (Figure 3A), the observed AI-aided MESE T2 protocol bias for Sys2 (pt3 and pt4) is relatively constant with histogram peaks mostly shifting to lower values by −7 ms to −5 ms, except for a +5 ms shift for Atr-mimic. In contrast, DL-aided MEEPI T2 protocol bias for Sys1 (pt1 and pt2) is markedly nonlinear with increasing negative shifts from −10 ms to −37 ms for nTZ and nPZ mimics and positive +14 ms for Atr-mimic. The comparative analysis of MEEPI protocols implemented alongside MESE on Sys2 (where AIsr reconstruction was compatible with both EPI and SE acquisition) confirmed similarity of nonlinear T2 bias trends with Sys1 (Supplementary Figure S3). These results reflect T2 bias dependence on SE versus EPI acquisition protocol variants for T2 mapping (rather than AI/DL reconstruction). Finite differences observed between protocol repeats on the same system (Figure 3A, Supplementary Figure S3) are within measurement uncertainties (error bars) and likely due to variations in acquisition parameters (e.g., TR, Table 1) and scanner room temperatures (Table 3). Compared to uncorrected values (Supplementary Table S3), bias correction minimizes differences to reference T2 values (Supplementary Table S4), most effectively for nTZ and nPZ mimics. It also reduces variability for phantom T2 values across protocols, as is evident from a 5 to 6-fold reduction in SD for T2 of nPZ and Atr mimics (Supplementary Table S4).
ADC(4b) model bias (Figure 3B) is consistent between AI/DL reconstruction protocols and system repeats, and manifests primarily as a shift to lower ADC versus reference DK values (similar to Supplementary Figure S1B with regular reconstruction). This result is highly suggestive of bias stemming from the ADC fit model rather than AI/DL-aided reconstruction. Minor variability between ADC protocol repeats is mainly due to temperature and EPI distortion (e.g., evident for phantom Sys1-pt1 (DLpf) protocol at ADC = 0.75 µm2/ms). Different biases would be observed for the ADC(2b) model (Supplementary Table S3), leading to increased variability of measured values across protocols. Except for Atr-mimic (p = 0.62), bias correction significantly (p < 0.015) reduced T2 and ADC parameter differences both between protocols and with respect to reference values for the phantom materials (Figure 3, blue, Supplementary Table S4), down to residual deviations due to temperature dependence of measured parameters.
The precision (histogram peak HW) and SNR are apparently similar to those of the DK phantom reference (Table 3, Figure 3B) for ADC(4b) of both Sys2-pt3 (AIsr) and Sys1-pt2 (DLpf) protocols (HW = 0.04–0.07 µm2/ms, with 3 mm slices). The precision is marginally better than the reference for Sys1-pt1 and Sys2-pt4 protocols (HW = 0.02–0.05 µm2/ms, with 4 mm slices), suggesting overall efficient AI/DL-aided denoising for EPI DWI acquisition. The SNR of both Sys1-DLpf (5TE-MEEPI) and Sys2-AIsr (5TE-MESE) T2 mapping protocols is lower than that of the reference (Table 3, Figure 3A). Phantom T2 precision is marginally better for Sys2 (HW = 2–6 ms) than for Sys1 protocols (HW = 4–12 ms), excepting the “Atr” mimic (long T2 peak). Lower SNR and large nonlinear bias are observed for the Sys1 T2 (MEEPI) protocol, apparently due to EPI distortion (compared to distortion-free MESE).
It is noteworthy that AI/DL-aided reconstruction for phantom scans mainly affected precision (HW) and SNR, while absolute bias was due to acquisition protocol (T2 MEEPI versus MESE and TE-range) or fit-model (ADC b-range).

3.3. Quantitative bpMRI Harmonization

The use of measured protocol biases (Figure 3, Table 3) for harmonization of lesion T2 and ADC of PCa patients scanned on different MR platforms is illustrated in Figure 4. The lesion SNR tends to be visually higher on ADC and T2 maps compared to individual b-value and TE weighted images (Figure 1), likely owing to multiple images used for parametric mapping. Bias correction implemented by subtracting the correction factors measured for the phantom is manifesting in shifts to higher values both for T2 and ADC lesion histograms (Figure 4, green). Larger shifts had to be applied for bias correction of ADC(4b) (Sys1-pt1: 0.38 µm2/ms & Sys2-pt3: 0.36 µm2/ms) and T2-MEEPI (Sys1-pt1: 30 ms) compared to T2-MESE (Sys2-pt3: 10 ms). The lesion T2 and ADC(4b) values harmonized to the DK reference across protocols and scan systems are listed in Table 4. The reference parameter ranges in the phantom (Figure 4, inserts) were apparently adequate to accommodate measured lesion values but would need to be expanded for normal (non-lesion) tissue bias correction (e.g., higher T2 and ADC values are needed for atrophy mimics).
After bias correction, the harmonized T2 and ADC values for lesion histograms (Table 4) can be quantitatively compared across patient scans independent of acquisition protocol. Similarly to phantom results, the main differences (Δ) in lesion T2 values before and after correction are associated with MEEPI (ΔT2 = 20–30 ms) versus MESE (ΔT2 = 9–10 ms) protocols, while for ADC(4b) bias is primarily determined by the fit model (ΔADC = 0.36–0.38 µm2/ms, Figure 5A). The bias correction results in equal T2 values for pt1 and pt3 lesions and confirms pt4 lesion as having relatively low ADC(DK) and T2 values, suggesting a higher likelihood of clinically significant PCa. The HW of lesion T2 (Table 4) for Sys2 (pt3 and pt4) is about half of Sys1 (pt1 and pt2), likely reflecting better SNR and lower distortion for MESE versus MEEPI acquisition. The HW for ADC(4b) is apparently correlated to the DWI slice width doubling for 3 mm versus 4 mm slices, both for Sys1 (DLpf) and Sys2 (AIsr) protocols. This is likely related to intrinsic limits of the corresponding DL/AI denoising models.
Phantom multiexponential diffusion materials allow evaluation of ADC dependence on b-value (Supplementary Figure S1). Consistent increase in ADC(DK) versus AIsr-ADC(2b) versus ADC(4b) is observed for phantom (Supplementary Figure S1, Tables S3 and S4) and PCa patient lesions (Figure 5B). The bias measured for mp-phantom (Figure 3B) closely represents differences between ADC models in vivo (Figure 5B, circles versus asterisks). The average difference between ADC(4b) and ADC(2b) fit models with measured lesion ADC(DK) is reduced 5–6-fold by correction, down to measurement uncertainty levels (from 0.42 and 0.21 µm2/ms to 0.07 and 0.04 µm2/ms). For studies that do not include high-b acquisition, ADC(2b) values derived in a lower-b-range would require harmonization with values derived using high-b ADC protocols (mean bias of 0.21 µm2/ms, Figure 5). Without bias correction, the use of low-b ADC(2b) to synthesize high-b DWI (Supplementary Figure S1A) would also artefactually decrease SNR for tissue with true kurtotic diffusion restriction, lowering the sensitivity of lesion detection. The HW for ADC(4b) lesion histograms is about half of ADC(2b) or ADC(DK). Thus, in practice, the 4b-fit may be preferred for the best achievable ADC SNR, and derived ADC(4b) values can be aligned with DK reference values by retrospective bias correction (as illustrated in this work). Note that proposed harmonization cannot be achieved with a mono-exponential diffusion references (e.g., PVP) that have no ADC dependence on b-value or with a reference tissue that has no or low kurtosis (e.g., body fluids).

4. Discussion

AI/DL accelerated acquisition [20,21] is desired to enable clinical implementation of quantitative bi-parametric MRI for improved specificity of clinically significant PCa risk stratification [6,8,9,14]. The goal of this study was to build a clinically viable validation workflow for AI/DL-aided reconstruction and denoising of prostate q-bpMRI and harmonize derived T2 and ADC values across different vendor platforms and acquisition protocols. Our study demonstrated the feasibility of flexible multi-system harmonization of q-bpMRI using AI/DL accelerated acquisitions by employing a multiparametric reference platform. This realistic-complexity platform allows direct quantitative testing of vendor-provided protocols with arbitrary settings for in vivo acquisition parameters with respect to the reference T2 and ADC values. The implemented workflow enabled assessment of individual bias contributions from acquisition protocols versus DL/AI-aided reconstruction versus parametric fit models. The performed phantom reference-based harmonization follows FDA guidelines for quantitative imaging devices [26] that are being adopted by quantitative imaging modalities [33,34]. The described method is generally more practical for prospective clinical application on a per-patient/per-protocol basis compared to traveling volunteer or reference tissue harmonization approaches [10,35].
Our study confirmed the benefits of AI/DL accelerated bpMRI for prostate imaging [20,21,32,36] that achieved two-fold acquisition time saving with acceptable diagnostic quality (comparable to SOC) on two clinical 3T scanners. Importantly, the tested vendor-provided AI/DL tools improved SNR or resolution for EPI-DWI without added bias and were compatible with EPI acquisition on both vendor systems. The qualitative tests for ADC and DWI confirmed higher efficiency of AI/DL-aided reconstruction for EPI-DWI denoising and of the AIsr method for improved resolution and reduced distortion, outperforming SOC. However, improved denoising was limited by slice thickness (to 4 mm), while the DLpf method appeared more susceptible to EPI distortion. For T2 mapping, the AIsr combination with MESE provided the best performance. These results are consistent with other prostate MRI studies that reported qualitative evaluations of AI/DL-aided reconstruction and denoising [20,21,22,32,36].
Much of the current research on multi-system harmonization of bi-parametric prostate MRI has been focused on downstream lesion segmentation and classification analysis in retrospective studies [37,38,39]. Harmonized quantitative metrics are a prerequisite for advanced predictive modeling using radiomics [40,41,42] and improved yield of multicenter clinical trials [14,24,27,37]. In the absence of reference standards and prospective acquisition protocol standardization, such studies largely rely on statistical and machine learning methods for retrospective data normalization [38,41,42]. Upstream elimination of large technical biases by reference standard harmonization for multiple acquisition conditions [26,33,34] improves generalization and reduces the numbers required for training downstream statistical harmonization models [37,38,41,42]. While QIBA recommendations for prospective standardization of acquisition protocol parameters [14,25,27,43] are being implemented in multicenter clinical trials to help minimize cross-platform biases in longitudinal studies (e.g., useful for treatment response assessment), their translation to SOC clinical practice remains challenging, particularly for diagnostic applications [14,24,27,43]. Cross-platform standardization of vendor-provided AI/DL-aided reconstruction based on proprietary algorithms and models [23] introduces an additional level of complexity. Regulatory acceptance of harmonized quantitative biomarkers in prostate cancer management necessitates validation of all components of the AI/DL-aided acquisition and analysis workflow [26,27]. The previous in vivo studies of AI/DL-accelerated acquisition protocols primarily conducted qualitative assessment for patients scanned on a single vendor platform [22,32,36] or utilized a single-parameter phantom that required a dedicated evaluation protocol [44].
The present prospective study implements the recommended harmonization approach [26,27,33,34] by combining qualitative assessment of AI/DL-aided q-bpMRI acquisition with quantitative bias measurement using a multiparametric reference phantom compatible with clinical protocols across different vendor systems. Multiparametric reference [28] allows objective testing of the SOC versus AI/DL reconstruction enhancements [23,26]. The implemented workflow enabled characterization of two distinct sources of biases in derived quantitative T2 and ADC parameters: (1) scan and reconstruction protocol settings and (2) fit model. Using the reference phantom, we measured and corrected the protocol-dependent biases in T2 and ADC values to bring them on a comparable scale for MRI visible lesions. Nonlinear bias was detected for T2-MEEPI versus nearly constant for T2-MESE, while ADC bias was due to b-range dependence in the presence of diffusion kurtosis. Importantly, our results suggest that reference tissue normalization based on body fluids [10] could be generally inadequate for prostate ADC and T2 across system protocols, when reference values have different dependence on acquisition parameters compared to PCa lesions due to tissue-dependent diffusion kurtosis or nonlinear T2 biases. Measurement accuracy for reference tissue (e.g., bladder [10] or muscle) also depends on anatomic coverage, which may be limited for reduced field-of-view acquisitions [31]. Demonstrated consistency of bias dependences in phantom and lesions, and for system protocol repeats, suggests that practical bias assessments and protocol optimizations are possible both prior to and after patient scans using mp-phantom for the ranges of target clinical acquisition parameters.
Our study had several limitations. The tests were performed on two systems with a small number of MR-visible low-grade lesions and a two-reader assessment. This may potentially increase the variability of qualitative scores for subjective evaluation, although we observed good inter-reader agreement and consistency with quantitative phantom SNR evaluation. Larger population, multi-system, multi-reader studies with correlation to clinical outcomes (including high and low PCa grades) are needed to ascertain the added benefit of the AI/DL-accelerated q-bpMRI protocols for PCa patient management [14,23,36]. The AI/DL parameters were not optimized, and default settings were used provided by vendor implementations for a realistic illustration of the clinical application. Further optimization is possible [20,21,31,36] to increase MESE T2 resolution, achieve compatibility with non-EPI acquisition, and optimize EPI-distortion correction.
Harmonization was based on the correction of additive bias for mean ROI metrics, which do not affect voxel-based image texture features and histograms used for radiomics analysis [40,41,42]. For such applications, the corrected T2 and ADC maps could be utilized to generate synthetic DWI and T2w images harmonized by using the same b-values and TEs. For more accurate bias measurement and correction, there is also a need to increase the reference range for ADC and T2 of mp-phantom to accommodate values for non-lesion tissue and perform temperature calibration studies for reference values within the scanner room temperature ranges [45,46]. These calibrations can be performed once and used prospectively to improve the precision of bias measurement for clinical protocols at ambient scanner room temperatures. Assessment of T2 biases due to varying protocol TR could also be improved by including phantom materials with tissue-mimicking T1 values. Finally, additional multi-system, multi-TE MESE and multi-b DWI in vivo studies are needed to confirm the consistency and accuracy of T2(TE) protocol biases and ADC b-range biases detected with the phantom.

5. Conclusions

The developed protocol enables harmonization of the vendor-provided AI/DL-accelerated acquisition options to establish credibility and streamline implementation of prostate q-bpMRI protocols in a multi-vendor clinical environment. AI/DL-aided reconstruction reduces acquisition time with improved SNR and resolution versus the SOC protocol. A multiparametric phantom allows assessment of biases directly for clinical scan protocols with arbitrary acquisition settings on a per-patient basis. The observed biases were due to acquisition parameters for T2-mapping and the b-range model fit for ADC-mapping. No added biases were detected from AI/DL reconstruction and denoising. The proposed workflow facilitates improved reproducibility and accuracy of AI/DL-aided quantitative ADC and T2 mapping for prostate patients in the SOC setting.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/s25185858/s1, Figure S1: (A) Example of b-range dependence for different ADC fit models (color-coded in the legend) for log-DWI signal, Sb (top), and corresponding parametric maps (bottom). (B) ADC b-range dependence reproduced in the layered diffusion kurtosis (DK) reference phantom (at scan temperature Ts). The top stacked plot shows ADC calibration histograms for three fit models (color-coded in the legend) generated from the parametric maps under the plot for the ROI (bin sizes: 0.03mm2/ms) marked on the kurtosis (K) map. Color-bar shows common ADC (and K) scale for mono-exponential fit models using all b-values (“4b”), 2b-values (“b100b800”) and diffusion kurtosis (“DK”). The kurtosis values [K] ± 0.05 are listed next to reference (“ref”) DK histogram peaks (green). Reference peak labels mark corresponding prostate tissue mimics (GS7: Gleason7, TZ: transition-zone, PZ: peripheral-zone, Atr: atrophy. Figure S2: The stacked plots compare histograms for AI/DL-aided acquisition protocols (Table 1, color-coded in the legends) versus calibrated reference (green) of T2 (A, bin size 3 ms) and ADC (B, bin size 0.03 μm2/ms) generated from the corresponding parametric maps that share the common scales (color-bars.) Phantom scan temperatures (ts ± 0.5 °C) are listed above the ADC maps (right). Figure S3: T2 bias comparison for AI-accelerated MEEPI versus MESE scan protocols for multi-parametric phantom. The mean T2 biases measured on Sys2 with respect to reference values are plotted (blue and cyan asterisks) for two repeated MEEPI scans (TE = [25, 55, 85, 115, 145] ms, TR = 6 s, Ts = 21 °C & 23.5 °C). Error-bars correspond to the phantom ROI histogram peak half-width (HW, see Figure S2). Solid green line marks zero bias/difference with error-bars indicating the reference (“Ref”) histogram HW (see Figure S1). Dashed lines are quadratic fits for the measured biases: for Sys2 AIsr T2-MEEPI with fit coefficients [c2, c1, c0] = [0.004, −0.95, 37.57], and for T2–MESE [c2, c1, c0] = [0.002, −0.314, 10.513] (data shown in Figure 3A), as well as for Sys1 DLpf-accelerated T2-MEEPI [c2, c1, c0] = [0.01, −2.26, 102.23] (Figure 3A). Table S1: Measured internal standard ADC values used to derive phantom scan temperature, Ts, using NIST calibration [https://data.nist.gov/od/id/mds2-2366 accessed on 3 August 2025]. Table S2: DK phantom reference values for repeated calibration (b = 0, 200, 500, 800, 1500, 2000, 2500 s/mm2; Ts = 21.0 ± 0.5 °C). Table S3. ROI mean T2 and ADC parameters for bp-phantom tissue mimics across different scan protocols. Table S4. ROI mean T2 and ADC parameters for bp-phantom tissue mimics corrected using bias-fit for different scan protocols with respect to DK reference values. References [28,30,46,47] are cited in the Supplementary Materials.

Author Contributions

Conceptualization, D.M., S.D.S. and T.L.C.; methodology, D.M., S.A.W., Y.J. and T.L.C.; scan-protocol implementation, D.M., S.D.S., J.R. and S.L.; data acquisition, J.R., S.L. and J.O.; software, D.M. and T.L.C.; validation, S.A.W., R.C. and T.L.C.; formal analysis, D.M., S.A.W., R.C., and S.D.S.; resources, D.M., S.D.S. and Y.J.; data curation, D.M., S.A.W., J.R., S.L., J.O. and T.L.C.; writing—original draft preparation D.M., S.D.S. and T.L.C.; writing—review and editing, all authors; visualization, D.M., S.A.W. and T.L.C.; supervision, D.M., S.D.S. and T.L.C.; project administration, D.M., Y.J. and J.R.; funding acquisition, internal funding to D.M. and Y.J. All authors have read and agreed to the published version of the manuscript.

Funding

Funding support from National Institute of Health Grants: 75N91021C00036, U01CA211205, R01CA190299, and the University of Michigan Basic Radiologic Sciences innovative research award.

Institutional Review Board Statement

The Institutional Review Board approved the study (IRB protocol # HUM 00175599, approved on 20 March 2024).

Informed Consent Statement

Informed consent was obtained from all subjects involved in the study.

Data Availability Statement

Essential data is contained in the article and Supplementary Figures.

Acknowledgments

We acknowledge the support of clinical research coordinator, Sarah Culver, for patient scheduling and consenting to the IRB for MR technology development. We are thankful to MRI technologist, Jesus E Fajardo, and vendor representatives, Nig Jin and Ramin Jafari, for their help with the initial implementation of WIP DL/AI protocols on clinical 3T systems.

Conflicts of Interest

S.D.S., T.L.C., and D.M. are co-inventors of diffusion kurtosis phantom IP, assigned to and managed by the University of Michigan.

Abbreviations

The following abbreviations are used in this manuscript:
ADCApparent diffusion coefficient
QIBAQuantitative Imaging Biomarker Alliance
NISTNational Institute of Standards
DWIDiffusion weighted imaging
T2wT2 weighted
DLpfDeep-learning partial-Fourier
AIsrArtificial-intelligence super-resolution
q-bpMRIQuantitative bi-parametric MRI
SOCStandard-of-care
MESEMulti-Echo Spin-Echo
MEEPIMulti-Echo echo-planar imaging
TEEcho-time
DKDiffusion kurtosis
HWHalf-width
Syssystem
ptpatient
PVPpolyvinylpyrrolidone

References

  1. Ferro, M.; Crocetto, F.; Bruzzese, D.; Imbriaco, M.; Fusco, F.; Longo, N.; Napolitano, L.; La Civita, E.; Cennamo, M.; Liotti, A.; et al. Prostate Health Index and Multiparametric MRI: Partners in Crime Fighting Overdiagnosis and Overtreatment in Prostate Cancer. Cancers 2021, 13, 4723. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  2. Schoots, I.G.; Barentsz, J.O.; Bittencourt, L.K.; Haider, M.A.; Macura, K.J.; Margolis, D.J.A.; Moore, C.M.; Oto, A.; Panebianco, V.; Siddiqui, M.M.; et al. PI-RADS Committee Position on MRI Without Contrast Medium in Biopsy-Naive Men With Suspected Prostate Cancer: Narrative Review. AJR Am. J. Roentgenol. 2021, 216, 3–19. [Google Scholar] [CrossRef] [PubMed]
  3. Shankar, P.R.; Maturen, K.E.; George, A.K.; Borza, T.; Ellimoottil, C.; Montgomery, J.S.; Wei, J.T.; Denton, B.T.; Davenport, M.S. Temporary Health Impact of Prostate MRI and Transrectal Prostate Biopsy in Active Surveillance Prostate Cancer Patients. J. Am. Coll. Radiol. 2019, 16, 1385–1392. [Google Scholar] [CrossRef] [PubMed]
  4. Purysko, A.S.; Rosenkrantz, A.B.; Barentsz, J.O.; Weinreb, J.C.; Macura, K.J. PI-RADS Version 2: A Pictorial Update. Radiographics 2016, 36, 1354–1372. [Google Scholar] [CrossRef] [PubMed]
  5. Barkovich, E.J.; Shankar, P.R.; Westphalen, A.C. A Systematic Review of the Existing Prostate Imaging Reporting and Data System Version 2 (PI-RADSv2) Literature and Subset Meta-Analysis of PI-RADSv2 Categories Stratified by Gleason Scores. AJR Am. J. Roentgenol. 2019, 212, 847–854. [Google Scholar] [CrossRef] [PubMed]
  6. Tavakoli, A.A.; Hielscher, T.; Badura, P.; Gortz, M.; Kuder, T.A.; Gnirs, R.; Schwab, C.; Hohenfellner, M.; Schlemmer, H.P.; Bonekamp, D. Contribution of Dynamic Contrast-enhanced and Diffusion MRI to PI-RADS for Detecting Clinically Significant Prostate Cancer. Radiology 2023, 306, 186–199. [Google Scholar] [CrossRef] [PubMed]
  7. Tan, C.H.; Hobbs, B.P.; Wei, W.; Kundra, V. Dynamic contrast-enhanced MRI for the detection of prostate cancer: Meta-analysis. AJR Am. J. Roentgenol. 2015, 204, W439–W448. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  8. Meyer, H.J.; Wienke, A.; Surov, A. Discrimination between clinical significant and insignificant prostate cancer with apparent diffusion coefficient—A systematic review and meta analysis. BMC Cancer 2020, 20, 482. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  9. Moraes, M.O.; Roman, D.H.H.; Copetti, J.; de S Santos, F.; Agra, A.; Noronha, J.A.P.; Carvalhal, G.; Neto, E.J.D.; Zanon, M.; Baldisserotto, M.; et al. Effects of the addition of quantitative apparent diffusion coefficient data on the diagnostic performance of the PI-RADS v2 scoring system to detect clinically significant prostate cancer. World J. Urol. 2020, 38, 981–991. [Google Scholar] [CrossRef] [PubMed]
  10. Nguyen, T.B.; Ushinsky, A.; Yang, A.; Nguyentat, M.; Fardin, S.; Uchio, E.; Lall, C.; Lee, T.; Houshyar, R. Utility of quantitative apparent diffusion coefficient measurements and normalized apparent diffusion coefficient ratios in the diagnosis of clinically significant peripheral zone prostate cancer. Br. J. Radiol. 2018, 91, 20180091. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  11. Hoang Dinh, A.; Souchon, R.; Melodelima, C.; Bratan, F.; Mege-Lechevallier, F.; Colombel, M.; Rouviere, O. Characterization of prostate cancer using T2 mapping at 3T: A multi-scanner study. Diagn. Interv. Imaging 2015, 96, 365–372. [Google Scholar] [CrossRef] [PubMed]
  12. Devine, W.; Giganti, F.; Johnston, E.W.; Sidhu, H.S.; Panagiotaki, E.; Punwani, S.; Alexander, D.C.; Atkinson, D. Simplified Luminal Water Imaging for the Detection of Prostate Cancer From Multiecho T(2) MR Images. J. Magn. Reson. Imaging 2019, 50, 910–917. [Google Scholar] [CrossRef] [PubMed]
  13. Panda, A.; O’Connor, G.; Lo, W.C.; Jiang, Y.; Margevicius, S.; Schluchter, M.; Ponsky, L.E.; Gulani, V. Targeted Biopsy Validation of Peripheral Zone Prostate Cancer Characterization With Magnetic Resonance Fingerprinting and Diffusion Mapping. Investig. Radiol. 2019, 54, 485–493. [Google Scholar] [CrossRef]
  14. Margolis, D.J.A.; Chatterjee, A.; deSouza, N.M.; Fedorov, A.; Fennessy, F.M.; Maier, S.E.; Obuchowski, N.; Punwani, S.; Purysko, A.; Rakow-Penner, R.; et al. Quantitative Prostate MRI, From the AJR Special Series on Quantitative Imaging. AJR Am. J. Roentgenol. 2024, 225, e2431715. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  15. Chatterjee, A.; Mercado, C.; Bourne, R.M.; Yousuf, A.; Hess, B.; Antic, T.; Eggener, S.; Oto, A.; Karczmar, G.S. Validation of Prostate Tissue Composition by Using Hybrid Multidimensional MRI: Correlation with Histologic Findings. Radiology 2022, 302, 368–377. [Google Scholar] [CrossRef]
  16. Hectors, S.J.; Semaan, S.; Song, C.; Lewis, S.; Haines, G.K.; Tewari, A.; Rastinehad, A.R.; Taouli, B. Advanced Diffusion-weighted Imaging Modeling for Prostate Cancer Characterization: Correlation with Quantitative Histopathologic Tumor Tissue Composition-A Hypothesis-generating Study. Radiology 2018, 286, 918–928. [Google Scholar] [CrossRef] [PubMed]
  17. Panagiotaki, E.; Chan, R.W.; Dikaios, N.; Ahmed, H.U.; O’Callaghan, J.; Freeman, A.; Atkinson, D.; Punwani, S.; Hawkes, D.J.; Alexander, D.C. Microstructural characterization of normal and malignant human prostate tissue with vascular, extracellular, and restricted diffusion for cytometry in tumours magnetic resonance imaging. Investig. Radiol. 2015, 50, 218–227. [Google Scholar] [CrossRef] [PubMed]
  18. Hurrell, S.L.; McGarry, S.D.; Kaczmarowski, A.; Iczkowski, K.A.; Jacobsohn, K.; Hohenwalter, M.D.; Hall, W.A.; See, W.A.; Banerjee, A.; Charles, D.K.; et al. Optimized b-value selection for the discrimination of prostate cancer grades, including the cribriform pattern, using diffusion weighted imaging. J. Med. Imaging 2018, 5, 011004. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  19. Simpkin, C.J.; Morgan, V.A.; Giles, S.L.; Riches, S.F.; Parker, C.; deSouza, N.M. Relationship between T2 relaxation and apparent diffusion coefficient in malignant and non-malignant prostate regions and the effect of peripheral zone fractional volume. Br. J. Radiol. 2013, 86, 20120469. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  20. Gassenmaier, S.; Afat, S.; Nickel, D.; Mostapha, M.; Herrmann, J.; Othman, A.E. Deep learning-accelerated T2-weighted imaging of the prostate: Reduction of acquisition time and improvement of image quality. Eur. J. Radiol. 2021, 137, 109600. [Google Scholar] [CrossRef] [PubMed]
  21. Ueda, T.; Ohno, Y.; Yamamoto, K.; Murayama, K.; Ikedo, M.; Yui, M.; Hanamatsu, S.; Tanaka, Y.; Obama, Y.; Ikeda, H.; et al. Deep Learning Reconstruction of Diffusion-weighted MRI Improves Image Quality for Prostatic Imaging. Radiology 2022, 303, 373–381. [Google Scholar] [CrossRef] [PubMed]
  22. van Lohuizen, Q.; Roest, C.; Simonis, F.F.J.; Fransen, S.J.; Kwee, T.C.; Yakar, D.; Huisman, H. Assessing deep learning reconstruction for faster prostate MRI: Visual vs. diagnostic performance metrics. Eur. Radiol. 2024, 34, 7364–7372. [Google Scholar] [CrossRef] [PubMed]
  23. Lin, D.J.; Johnson, P.M.; Knoll, F.; Lui, Y.W. Artificial Intelligence for MR Image Reconstruction: An Overview for Clinicians. J. Magn. Reson. Imaging 2021, 53, 1015–1028. [Google Scholar] [CrossRef]
  24. Vergara, D.; Armato, S.G.; Hadjiiski, L., 3rd; Drukker, K. Best Practices for Artificial Intelligence and Machine Learning for Computer-Aided Diagnosis in Medical Imaging. J. Am. Coll. Radiol. 2024, 21, 341–343. [Google Scholar] [CrossRef]
  25. Obuchowski, N.A.; Huang, E.; deSouza, N.M.; Raunig, D.; Delfino, J.; Buckler, A.; Hatt, C.; Wang, X.; Moskowitz, C.; Guimaraes, A.; et al. A Framework for Evaluating the Technical Performance of Multiparameter Quantitative Imaging Biomarkers (mp-QIBs). Acad. Radiol. 2023, 30, 147–158. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  26. US FDA Center for Devices and Radiological Health. Technical Performance Assessment of Quantitative Imaging in Radiological Device Premarket Submissions; FDA: Silver Spring, MA, USA, 2022. Available online: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/technical-performance-assessment-quantitative-imaging-radiological-device-premarket-submissions (accessed on 3 August 2025).
  27. Boss, M.A.; Malyarenko, D.; Partridge, S.; Obuchowski, N.; Shukla-Dave, A.; Winfield, J.M.; Fuller, C.D.; Miller, K.; Mishra, V.; Ohliger, M.; et al. The QIBA Profile for Diffusion-Weighted MRI: Apparent Diffusion Coefficient as a Quantitative Imaging Biomarker. Radiology 2024, 313, e233055. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  28. Malyarenko, D.; Ono, S.; Lynch, T.J.E.; Swanson, S.D. Technical note: Hydrogel-based mimics of prostate cancer with matched relaxation, diffusion and kurtosis for validating multi-parametric MRI. Med. Phys. 2024, 51, 3590–3596. [Google Scholar]
  29. Dieckmeyer, M.; Ruschke, S.; Eggers, H.; Kooijman, H.; Rummeny, E.J.; Kirschke, J.S.; Baum, T.; Karampinos, D.C. ADC Quantification of the Vertebral Bone Marrow Water Component: Removing the Confounding Effect of Residual Fat. Magn. Reson. Med. 2017, 78, 1432–1441. [Google Scholar]
  30. Malyarenko, D.; Swanson, S.D.; Richardson, J.; Lowe, S.; O’Connor, J.; Fajardo, J.E.; Jiang, Y.; Wells, S.; Chenevert, T.L. Harmonization of AI/DL accelerated quantitative bi-parametric prostate MRI: Demonstration in multi-parametric phantom and patients. In Proceedings of the 2025 ISMRM, Honolulu, HI, USA, 10–15 May 2025; p. 103. [Google Scholar]
  31. Lawrence, E.M.; Zhang, Y.; Starekova, J.; Wang, Z.; Pirasteh, A.; Wells, S.A.; Hernando, D. Reduced field-of-view and multi-shot DWI acquisition techniques: Prospective evaluation of image quality and distortion reduction in prostate cancer imaging. Magn. Reson. Imaging 2022, 93, 108–114. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  32. Bischoff, L.M.; Peeters, J.M.; Weinhold, L.; Krausewitz, P.; Ellinger, J.; Katemann, C.; Isaak, A.; Weber, O.M.; Kuetting, D.; Attenberger, U.; et al. Deep Learning Super-Resolution Reconstruction for Fast and Motion-Robust T2-weighted Prostate MRI. Radiology 2023, 308, e230427. [Google Scholar] [CrossRef] [PubMed]
  33. Panetta, J.V.; Daube-Witherspoon, M.E.; Karp, J.S. Validation of phantom-based harmonization for patient harmonization. Med. Phys. 2017, 44, 3534–3544. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  34. Kim, H.; Thomas, J.V.; Nix, J.W.; Gordetsky, J.B.; Li, Y.; Rais-Bahrami, S. Portable Perfusion Phantom Offers Quantitative Dynamic Contrast-Enhanced Magnetic Resonance Imaging for Accurate Prostate Cancer Grade Stratification: A Pilot Study. Acad. Radiol. 2021, 28, 405–413. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  35. Saito, Y.; Kamagata, K.; Andica, C.; Maikusa, N.; Uchida, W.; Takabayashi, K.; Yoshida, S.; Hagiwara, A.; Fujita, S.; Akashi, T.; et al. Traveling Subject-Informed Harmonization Increases Reliability of Brain Diffusion Tensor and Neurite Mapping. Aging. Dis. 2023, 15, 2770–2785. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  36. Johnson, P.M.; Tong, A.; Donthireddy, A.; Melamud, K.; Petrocelli, R.; Smereka, P.; Qian, K.; Keerthivasan, M.B.; Chandarana, H.; Knoll, F. Deep Learning Reconstruction Enables Highly Accelerated Biparametric MR Imaging of the Prostate. J. Magn. Reson. Imaging 2022, 56, 184–195. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  37. Saha, A.; Bosma, J.S.; Twilt, J.J.; van Ginneken, B.; Bjartell, A.; Padhani, A.R.; Bonekamp, D.; Villeirs, G.; Salomon, G.; Giannarini, G.; et al. PI-CAI consortium. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): An international, paired, non-inferiority, confirmatory study. Lancet Oncol. 2024, 25, 879–887. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  38. Giganti, F.; Moreira da Silva, N.; Yeung, M.; Davies, L.; Frary, A.; Ferrer Rodriguez, M.; Sushentsev, N.; Ashley, N.; Andreou, A.; Bradley, A.; et al. AI-powered prostate cancer detection: A multi-centre, multi-scanner validation study. Eur. Radiol. 2025, 35, 4915–4924. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  39. Sunoqrot, M.R.S.; Saha, A.; Hosseinzadeh, M.; Elschot, M.; Huisman, H. Artificial intelligence for prostate MRI: Open datasets, available applications, and grand challenges. Eur. Radiol. Exp. 2022, 6, 35. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  40. Telecan, T.; Chiorean, A.; Sipos-Lascu, R.; Caraiani, C.; Boca, B.; Hendea, R.M.; Buliga, T.; Andras, I.; Crisan, N.; Lupsor-Platon, M. ISUP Grade Prediction of Prostate Nodules on T2WI Acquisitions Using Clinical Features, Textural Parameters and Machine Learning-Based Algorithms. Cancers 2025, 17, 2035. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  41. Castaldo, R.; Brancato, V.; Cavaliere, C.; Trama, F.; Illiano, E.; Costantini, E.; Ragozzino, A.; Salvatore, M.; Nicolai, E.; Franzese, M. A Framework of Analysis to Facilitate the Harmonization of Multicenter Radiomic Features in Prostate Cancer. J. Clin. Med. 2022, 12, 140. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  42. Hu, F.; Chen, A.A.; Horng, H.; Bashyam, V.; Davatzikos, C.; Alexander-Bloch, A.; Li, M.; Shou, H.; Satterthwaite, T.D.; Yu, M.; et al. Image harmonization: A review of statistical and deep learning methods for removing batch effects and evaluation metrics for effective harmonization. Neuroimage 2023, 274, 120125. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  43. Boss, M.A.; Snyder, B.S.; Kim, E.; Flamini, D.; Englander, S.; Sundaram, K.M.; Gumpeni, N.; Palmer, S.L.; Choi, H.; Froemming, A.T.; et al. Repeatability and Reproducibility Assessment of the Apparent Diffusion Coefficient in the Prostate: A Trial of the ECOG-ACRIN Research Group (ACRIN 6701). J. Magn. Reson. Imaging 2022, 56, 668–679. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  44. Lemainque, T.; Yoneyama, M.; Morsch, C.; Iordanishvili, E.; Barabasch, A.; Schulze-Hagen, M.; Peeters, J.M.; Kuhl, C.; Zhang, S. Reduction of ADC bias in diffusion MRI with deep learning-based acceleration: A phantom validation study at 3.0 T. Magn. Reson. Imaging 2024, 110, 96–103. [Google Scholar] [CrossRef] [PubMed]
  45. Keenan, K.E.; Peskin, A.P.; Wilmes, L.J.; Aliu, S.O.; Jones, E.F.; Li, W.; Kornak, J.; Newitt, D.C.; Hylton, N.M. Variability and bias assessment in breast ADC measurement across multiple systems. J. Magn. Reson. Imaging 2016, 44, 846–855. [Google Scholar] [CrossRef]
  46. Amouzandeh, G.; Chenevert, T.L.; Swanson, S.D.; Ross, B.D.; Malyarenko, D.I. Technical note: Temperature and concentration dependence of water diffusion in polyvinylpyrrolidone solutions. Med. Phys. 2022, 49, 3325–3332. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  47. Holz, M.; Heil, S.R.; Sacco, A. Temperature-dependent self-diffusion coefficients of water and six selected molecular 16 liquids for calibration in accurate 1H NMR PFG measurements. Phys. Chem. Chem. Phys. 2000, 2, 4740–4742. [Google Scholar] [CrossRef]
Sensors 25 05858 g001
Figure 1. Gray-scale images qualitatively compare examples of standard of care (SOC) bi-parametric (bp)MRI versus deep-learning partial-Fourier (DLpf) and artificial-intelligence super-resolution (AIsr) from two different vendor 3T systems (Sys1 and Sys2) in two patients (pt1 and pt3) with PIRADS 4 lesions (arrows) in the right anterior mid-gland (pt1, left) and left posterior apex (pt3, right) peripheral zone.
Figure 1. Gray-scale images qualitatively compare examples of standard of care (SOC) bi-parametric (bp)MRI versus deep-learning partial-Fourier (DLpf) and artificial-intelligence super-resolution (AIsr) from two different vendor 3T systems (Sys1 and Sys2) in two patients (pt1 and pt3) with PIRADS 4 lesions (arrows) in the right anterior mid-gland (pt1, left) and left posterior apex (pt3, right) peripheral zone.
Sensors 25 05858 g001
Sensors 25 05858 g002
Figure 2. Different sensitivity to susceptibility distortions for standard of care (SOC1/SOC2) versus DL/AI accelerated bi-parametric (bp)MRI on two 3T systems (Sys1 and Sys2) for two patients (pt) with peripheral zone PIRADS 5 lesions (arrows) in the right apex (pt2) and left apex (pt4).
Figure 2. Different sensitivity to susceptibility distortions for standard of care (SOC1/SOC2) versus DL/AI accelerated bi-parametric (bp)MRI on two 3T systems (Sys1 and Sys2) for two patients (pt) with peripheral zone PIRADS 5 lesions (arrows) in the right apex (pt2) and left apex (pt4).
Sensors 25 05858 g002
Sensors 25 05858 g003
Figure 3. Quantitative bias assessment for multiparametric phantom scanned with four DL/AI accelerated bpMRI protocols (“Sys1-pt1”, “Sys1-pt2”, “Sys2-pt3”, “Sys2-pt4”, color-coded in the legends) used for patients (pt) with MRI-visible lesions (Table 1). The plots summarize mean bias (asterisks) versus reference values for T2 (A) and ADC (B). Error bars correspond to the phantom ROI histogram peak half-width (HW, see Supplementary Figure S2). Dashed lines are fits for the measured biases: constant for ADC [Sys1-pt1,2 and Sys2-pt3,4] = [0.38 and 0.36] µm2/ms; and quadratic for T2 (Sys1-MEEPI and Sys2-MESE bias fit coefficients listed in Supplementary Figure S3). Solid blue lines indicate residual parameter differences between protocols after corresponding bias correction (“Cor”). Solid green line marks zero bias/difference with error-bars indicating the reference (“Ref”) histogram HW (see Supplementary Figure S1).
Figure 3. Quantitative bias assessment for multiparametric phantom scanned with four DL/AI accelerated bpMRI protocols (“Sys1-pt1”, “Sys1-pt2”, “Sys2-pt3”, “Sys2-pt4”, color-coded in the legends) used for patients (pt) with MRI-visible lesions (Table 1). The plots summarize mean bias (asterisks) versus reference values for T2 (A) and ADC (B). Error bars correspond to the phantom ROI histogram peak half-width (HW, see Supplementary Figure S2). Dashed lines are fits for the measured biases: constant for ADC [Sys1-pt1,2 and Sys2-pt3,4] = [0.38 and 0.36] µm2/ms; and quadratic for T2 (Sys1-MEEPI and Sys2-MESE bias fit coefficients listed in Supplementary Figure S3). Solid blue lines indicate residual parameter differences between protocols after corresponding bias correction (“Cor”). Solid green line marks zero bias/difference with error-bars indicating the reference (“Ref”) histogram HW (see Supplementary Figure S1).
Sensors 25 05858 g003
Sensors 25 05858 g004
Figure 4. Bi-parametric Sys1-DLpf (top row) and Sys2-AIsr (bottom row) maps (color) and lesion histograms for T2 and ADC(4b) (bin sizes: 5 ms and 0.05 µm2/ms) generated for two patients (pt1 and pt3) with PIRADS 4 lesions (arrows) scanned on different clinical 3T MRIs (Sys1, Sys2). Color bars indicate quantitative parameter scales. The inserts show corresponding maps for the multiparametric phantom used to assess protocol biases (Figure 3) that are corrected (“Cor”) to harmonize T2 and ADC values for lesion histograms (right).
Figure 4. Bi-parametric Sys1-DLpf (top row) and Sys2-AIsr (bottom row) maps (color) and lesion histograms for T2 and ADC(4b) (bin sizes: 5 ms and 0.05 µm2/ms) generated for two patients (pt1 and pt3) with PIRADS 4 lesions (arrows) scanned on different clinical 3T MRIs (Sys1, Sys2). Color bars indicate quantitative parameter scales. The inserts show corresponding maps for the multiparametric phantom used to assess protocol biases (Figure 3) that are corrected (“Cor”) to harmonize T2 and ADC values for lesion histograms (right).
Sensors 25 05858 g004
Sensors 25 05858 g005
Figure 5. (A) Average lesion T2 versus ADC values are plotted before (blue asterisks) and after bias correction (green circles). The error bars correspond to the T2 histogram half-width (HW). (B) Efficiency of ADC model harmonization (4b, 2b versus DK) is illustrated for four lesions by alignment of measured (green asterisks) and corrected (blue and black circles) ADC values. The data are color-coded in the legend. The error bars show the representative ADC histogram HW for the first lesion. The lines in (A,B) connect the measured data (symbols) for visual guidance.
Figure 5. (A) Average lesion T2 versus ADC values are plotted before (blue asterisks) and after bias correction (green circles). The error bars correspond to the T2 histogram half-width (HW). (B) Efficiency of ADC model harmonization (4b, 2b versus DK) is illustrated for four lesions by alignment of measured (green asterisks) and corrected (blue and black circles) ADC values. The data are color-coded in the legend. The error bars show the representative ADC histogram HW for the first lesion. The lines in (A,B) connect the measured data (symbols) for visual guidance.
Sensors 25 05858 g005
Table 1. Essential acquisition protocol parameters.
Table 1. Essential acquisition protocol parameters.
ProtocolTE (ms)TR (s)b[nav] (s/mm2)Acquired Voxel (mm3)Scan Duration (min)
Sys1:
SOC1-T2w107* 9.2/4.8NA0.5 × 0.6 × 36:00
SOC1-DWI914.4/4.80[1], 100[1], 800[3], 1600[20]1.75 × 1.75 × 48:50
DLpf-T2 (MEEPI)40, 70, 100, 130, 1605.5/6.8NA2 × 2 × 3
(off-line map)		2:10
DLpf-ADC805.50[1], 100[1], 800[2], 1600[8]2 × 2 × 4/33:30
Sys2:
SOC2-T2w1104.4NA0.4 × 0.7 × 35:00
SOC2-DWI777.20[2], 100[2], 800[4], 1600[16]2.2 × 2.3 × 48:20
AIsr-T2 (MESE)25, 65, 105, 145, 1858/12.2NA2 × 2.3 × 31:45
AIsr-ADC773.9/5.40[1], 100[1], 800[2], 1600[8]2 × 2 × 3/44:00
* Parameters in bold changed between patient scans; TE: echo time; TR: repeat time; [nav]: number of averages per b-value; Sys1: Siemens Vida; Sys2: Philips Ingenia; SOC: standard of care; DLpf: deep-learning partial-Fourier; AIsr: artificial-intelligence super-resolution; MEEPI: multi-echo echo-planer imaging; MESE: multi-echo spin-echo.
Table 2. Likert-like scores (rater 1/rater 2) for qualitative comparison of SOC versus AI/DL-accelerated bpMRI for patients with MR-visible lesions.
Table 2. Likert-like scores (rater 1/rater 2) for qualitative comparison of SOC versus AI/DL-accelerated bpMRI for patients with MR-visible lesions.
ImageEval. CriteriaSys1-pt1
(PIRADS 4)		Sys1-pt2
(PIRADS 5)		Sys2-pt3
(PIRADS 4)		Sys2-pt4
(PIRADS 5)
ADCDx quality3/32/33/43/3
distortion3/32/24/34/4
resolution3/22/24/44/4
SNR4/32/23/44/4
median3/32/23.5/44/4
DWIb1600Dx quality3/32/23/34/3
distortion3/32/24/34/4
resolution2/22/24/33/4
SNR4/42/22/34/4
median3/32/23.5/34/4
T2wDx quality2/21/13/34/3
distortion2/31/23/33/3
resolution2/22/22/23/2
SNR3/22/24/44/4
median2/21.5/23/33.5/3
Sys-pt: system—patient; ADC: apparent diffusion coefficient; SNR: signal-to-noise ratio; DWIb1600: diffusion weighted image for b = 1600 s/mm2; T2w: T2 weighted image; Dx: diagnostic.
Table 3. T2 and ADC parameters (histogram mean and half-width) measured for bp-phantom tissue mimics using patient-specific AI/DL-aided Sys1 and Sys2 protocols and reference scans.
Table 3. T2 and ADC parameters (histogram mean and half-width) measured for bp-phantom tissue mimics using patient-specific AI/DL-aided Sys1 and Sys2 protocols and reference scans.
Protocol
(Ts ± 0.5 °C)		Parameter
Mean [HW]		GS7nTZnPZAtr
DKref (22.0)ADC ± 0.015 (mm2/ms)1.06 [0.06]1.33 [0.06]1.42 [0.04]1.72 [0.04]
Sys1-pt1 (21.4)0.64 [0.04]0.91 [0.04]1.03 [0.02]1.33 [0.02]
Sys1-pt2 (21.1)0.77 [0.05]0.92 [0.06]1.07 [0.06]1.31 [0.04]
Sys2-pt3 (21.0)0.79 [0.07]0.88 [0.06]1.03 [0.04]1.36 [0.04]
Sys2-pt4 (23.5)0.83 [0.05]0.92 [0.04]1.07 [0.03]1.43 [0.03]
T2ref (22.0)T2 ± 1.5 (ms)71 [4]89 [2]137 [3]173 [8]
Sys1-pt1 (21.4)64 [4]70 [8]109 [7]187 [7]
Sys1-pt2 (21.1)73 [12]79 [6]100 [6]166 [8]
Sys2-pt3 (21.0)67 [2]79 [4]130 [5]169 [10]
Sys2-pt4 (23.5)73 [4]82 [5]139 [6]178 [9]
Ts: phantom scan temperature; HW: half-width; GS7: Gleason7, TZ: transition-zone, PZ: peripheral-zone, Atr: atrophy tissue mimics; DKref: diffusion kurtosis reference; Sys-pt: system—patient protocol.
Table 4. Measured and corrected T2 and ADC (mean and half-width) for lesion histograms.
Table 4. Measured and corrected T2 and ADC (mean and half-width) for lesion histograms.
Parameter
Mean [HW]		Sys1-pt1Sys1-pt2Sys2-pt3Sys2-pt4
T2meas ± 3 (ms)90 [25]76 [20]110 [10]82 [16]
T2cor ± 3 (ms)120 [25]96 [20]120 [10]91 [16]
ADC(4b)meas ± 0.03 (mm2/ms)0.93 [0.09]1.01 [0.18]0.75 [0.24]0.56 [0.12]
ADC(4b)cor ± 0.03 (mm2/ms)1.31 [0.09]1.39 [0.18]1.11 [0.24]0.92 [0.12]
ADC(2b)meas ± 0.03 (mm2/ms)1.18 [0.16]1.16 [0.31]0.98 [0.36]0.77 [0.23]
ADC(2b)cor ± 0.03 (mm2/ms)1.37 [0.16]1.35 [0.31]1.16 [0.36]0.95 [0.23]
ADC(DK)meas ± 0.03 (mm2/ms)1.45 [0.2]1.36 [0.44]1.2 [0.38]0.91 [0.24]
HW: half-width; Sys-pt: system—patient; meas: measured; cor: corrected; ADC(4b): four b-value ADC fit; ADC(2b): two b-value ADC fit; ADC(DK): diffusion kurtosis ADC fit.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Malyarenko, D.; Swanson, S.D.; Richardson, J.; Lowe, S.; O’Connor, J.; Jiang, Y.; Chahine, R.; Wells, S.A.; Chenevert, T.L. Cross-Scanner Harmonization of AI/DL Accelerated Quantitative Bi-Parametric Prostate MRI. Sensors 2025, 25, 5858. https://doi.org/10.3390/s25185858

AMA Style

Malyarenko D, Swanson SD, Richardson J, Lowe S, O’Connor J, Jiang Y, Chahine R, Wells SA, Chenevert TL. Cross-Scanner Harmonization of AI/DL Accelerated Quantitative Bi-Parametric Prostate MRI. Sensors. 2025; 25(18):5858. https://doi.org/10.3390/s25185858

Chicago/Turabian Style

Malyarenko, Dariya, Scott D. Swanson, Jacob Richardson, Suzan Lowe, James O’Connor, Yun Jiang, Reve Chahine, Shane A. Wells, and Thomas L. Chenevert. 2025. "Cross-Scanner Harmonization of AI/DL Accelerated Quantitative Bi-Parametric Prostate MRI" Sensors 25, no. 18: 5858. https://doi.org/10.3390/s25185858

APA Style

Malyarenko, D., Swanson, S. D., Richardson, J., Lowe, S., O’Connor, J., Jiang, Y., Chahine, R., Wells, S. A., & Chenevert, T. L. (2025). Cross-Scanner Harmonization of AI/DL Accelerated Quantitative Bi-Parametric Prostate MRI. Sensors, 25(18), 5858. https://doi.org/10.3390/s25185858

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop