Predicting ADC Map Quality from T2-Weighted MRI: A Deep Learning Approach for Early Quality Assessment to Assist Point-of-Care

Abstract

Purpose:

Poor quality prostate MRI images compromise diagnostic accuracy, with diffusion-weighted imaging and the resulting apparent diffusion coefficient (ADC) maps being particularly vulnerable. These maps are critical for prostate cancer diagnosis, yet current methods relying on standardizing technical parameters fail to consistently ensure image quality. We propose a novel deep learning approach to predict low-quality ADC maps using T2-weighted (T2W) images, enabling real-time corrective interventions during imaging.

Materials and Methods:

A multi-site dataset of T2W images and ADC maps from 486 patients, spanning 62 external clinics and in-house imaging, was retrospectively analyzed. A neural network was trained to classify ADC map quality as “diagnostic” or “non-diagnostic” based solely on T2W images. Rectal cross-sectional area measurements were evaluated as an interpretable metric for susceptibility-induced distortions.

Results:

Analysis revealed limited correlation between individual acquisition parameters and image quality, with horizontal phase encoding significant for T2 imaging (p<0.001, AUC=0.6735) and vertical resolution for ADC maps (p=0.006, AUC=0.6348). By contrast, the neural network achieved robust performance for ADC map quality prediction from T2 images, with 83% sensitivity and 90% negative predictive value in multicenter validation, comparable to single-site models using ADC maps directly. Remarkably, it generalized well to unseen in-house data (94±2% accuracy). Rectal cross-sectional area correlated with ADC quality (AUC = 0.65), offering a simple, interpretable metric.

Conclusion:

The probability of low quality, uninterpretable ADC maps can be inferred early in the imaging process by a neural network approach, allowing corrective action to be employed.

Introduction

Accurate diagnosis of localized prostate cancer relies heavily on high-quality multi-parametric MRI (mpMRI). mpMRI has a high negative predictive value and acceptable positive predictive value when performed properly and has thus, become the initial step in the diagnostic pathway in patients with elevated prostate specific antigen (PSA) levels [1]. Because MRI is commonly used to guide prostate biopsies, poor quality images heavily influence all downstream events. Suboptimal mpMRI image quality can lead to delayed diagnosis, unnecessary biopsies, and misclassification of tumors, ultimately compromising patient care. A 2023 multicenter European study revealed that low-quality images were over three times more likely to be upgraded to a higher-grade, more dangerous status after biopsy [2]. 40% of images in this study fell into the low-quality category, indicating a significant number of potentially cancer harboring lesions may remain undetected by mpMRI due to image quality issues [2]. In the best case, a clinically nondiagnostic MRI requires another visit for repeat imaging, which is a misallocation of scarce resources. In the worst case, a suboptimal image misses the timely diagnosis of a potentially fatal cancer.

Efforts to enhance prostate image quality can be divided into quality control measures, which focus on establishing processes that favor high image quality, and quality assessment, which aims to identify low-quality images early in the diagnostic process [1]. As an example of a quality control initiative, the PI-RADS committee and the later PI-QUAL standard sought to standardize imaging protocols by setting minimum technical requirements for MRI acquisition, aiming to reduce variability between imaging sites [3, 4]. The success of these initiatives has been controversial, with some studies indicating continuous improvement as PI-RADSv2 technical requirements are met, [5] while others suggest a weak correlation between adherence to these standards and image quality [6]. While quality control methods are undoubtedly needed to meet quality standards [7, 8], quality assessment will remain a vital part of the imaging assessment pipeline [9].

Quality assessment can theoretically be performed manually by the radiologist during image acquisition; however, normal hospital operations rarely allow for this ideal scenario. In standard clinical practice, radiologists primarily interpret images after acquisition is complete rather than at the scanner itself. Automatic quality assessment by artificial intelligence can potentially alert the technologist to methodological problems earlier in the scanning process, allowing them to address these issues before the patient leaves the imaging facility without necessarily engaging the radiologist [10–12]. Automatic quality assessment currently suffers from two issues that may limit its clinical effectiveness. Although high accuracy in binary quality classification of T2 and ADC images has been achieved with deep learning in single-site settings where techniques are standardized [10], replicating this accuracy across multiple sites has proved challenging [12]. The ADC map is perhaps the single most valuable sequence for prostate cancer detection. However, it is particularly vulnerable to quality degradation due to the EPI sequence’s susceptibility to magnetic distortion, which frequently affects the posterior prostate region where the majority of clinically significant lesions are located. Current methods require waiting until after ADC map acquisition for quality assessment, which is inefficient since this sequence is the most time-consuming component of the mpMRI protocol and occurs late in the examination. This delayed assessment wastes valuable scanner time and extends patient appointments unnecessarily when poor-quality images require rescanning. In this study, we evaluated the predictive value of standardized technical parameters, finding no single acquisition parameter consistently predicted image quality. Given these constraints, we developed a deep learning approach using a multi-site training corpus which evaluates T2 images to predict the future quality of the ADC maps, allowing corrective actions before lengthy DWI sequences.

Methods

Study Population

The training data set consisted of mpMRI images from 486 patients (mean age 64.0 ± 7.3 SD, mean weight 85.3 kg ± 15.4 SD) imaged first at one of 62 different institutions before being subsequently referred to our facility for prostate imaging [13]. The external sites included both academic and private centers and no attempt was made to standardize the imaging among different sites. To ensure anonymity, all DICOM images underwent a two-step anonymization process: automated removal of header metadata followed by manual verification to ensure the absence of protected health information in both file names and image content.

Image quality was assessed by an experienced radiologist with over 15 years of experience in reading mpMRI of the prostate. This observer made note of both technical distortions (e.g., motion, artifacts, noise, and aliasing) and perceptual issues (e.g., blurred prostate capsule or zones, unclear external urethral sphincter, or excessive rectal gas). The quality evaluation criteria was based on PI-QUALv2.[14] Each image set received a 1-3 rating: 1 for poor quality (significantly hindering diagnosis), 2 for adequate quality (diagnostically usable despite minor distortions), and 3 for high quality (no distortions) (Fig. 1). To simplify analysis, the “high quality” and “adequate” categories were combined into a single “diagnostic” category while “poor quality” was considered “non-diagnostic”.

Figure 1:

Quality assessment of prostate MRI across institutions. (A) Representative axial T2-weighted images and corresponding ADC maps demonstrating the three-point quality scoring system (QC=3: optimal quality, QC=2: adequate quality, QC=1: non-diagnostic quality). (B) Venn diagrams showing the distribution of image quality between in house and outside institutions for both T2 and ADC sequences. (C) Correlation matrix between T2 and ADC quality scores for the same patient within and across institutions

Neural Network Building

To develop our predictive model, we implemented a multi-scale deep learning approach analyzing 14 consecutive axial slices centered on the imaging isocenter, which was assumed to be centered on the prostate midgland. Images were divided into three anatomically relevant, overlapping regions (bladder, prostate, and rectum) (Fig. 2B). The validation process employed a rigorous multi-stage approach to ensure unbiased model evaluation (Fig. 2B). First, 20% of the external patients were randomly held out as a final validation set. The remaining 80% of external patients underwent stratified 5-fold cross-validation, which maintains the same proportion of diagnostic and non-diagnostic cases in each fold. In each cross-validation iteration, four folds (80% of the remaining external data) combined with all internal patient data formed the training set, while the fifth fold of external patients served as the test set. This approach ensured the model was trained on a mix of internal and external data but evaluated exclusively on external data to assess generalizability across different imaging sites. A logistic regression model as described below was then trained on the test set predictions using 5-fold cross-validation for parameter optimization. The final performance metrics were obtained by applying this regression model to the completely independent validation set, using a classification threshold of 0.5.

Figure 2:

Schematic overview of the multi-stage deep learning approach for prostate MRI quality assessment. (A) Representative T2-weighted axial slice showing the three anatomically relevant regions (bladder, prostate, and rectum) used for analysis. (B) Workflow diagram illustrating the hierarchical integration of neural networks. Each anatomical region is processed by both Inception and ConvNext networks, with predictions combined through L2 regularized logistic regression. The model incorporates scores from other modalities and models to generate T2 predictions and final quality assessments, with early stopping mechanisms at each neural network stage.

For each region and modality (T2 and ADC), we trained two complementary neural networks: Google Inception [15, 16] was selected for its multi-scale perception through inception modules, and ConvNext [17] was selected for it transformer-like performance in a CNN architecture. Both networks were trained using CORN (Consistent Rank Logits) ordinal regression to predict quality scores on a three-point scale [18]. The networks were trained with adaptive sharpness-aware minimization (ASAM) [19] using a cosine annealing learning rate schedule (initial lr=5e-5, minimum lr=1e-6) and L2 regularization (weight decay=1e-2) with the Adam optimizer [20]. To address class imbalance [21], we implemented a weighted loss function where weights were inversely proportional to class frequency and multiplied the weight of the lowest quality category by 5 to enhance sensitivity for problematic scans. Data augmentation included random horizontal flipping, ±10° rotations, contrast/brightness adjustments, and cutout regularization [22] with a dropout rate of 0.90. Images were preprocessed using CLAHE (Contrast Limited Adaptive Histogram Equalization) [23] with a clip limit of 2.0 and 8x8 tile size, then center-cropped to 164 pixels and resized to 299x299. Early stopping based on maximum AUC in the test set and adaptive sharpness aware minimization (ASAM) was used during training to prevent overfitting [19].

The final model employed a hierarchical regression approach to combine predictions across networks, regions, and modalities. For each of the three anatomical regions, the three-point quality scores were collapsed into a binary classification (non-diagnostic vs diagnostic) before predictions from both Inception and ConvNext networks were combined using L2 regularized logistic regression (liblinear solver, balanced class weights, 5-fold cross-validation for parameter selection). The regional scores were then merged to create modality-specific predictions using another L2 regularized logistic regression model. Finally, T2 and ADC modality scores were combined through regularized regression to produce a single quality assessment score. To ensure unbiased performance assessment, the regression models were trained on the test set data (which was independent from the neural network training set) and evaluated on a separate validation set with no patient overlap as described above.

Rectal Cross-Sectional Area Analysis

While the neural network achieved robust predictive performance, its ‘black box’ nature makes it challenging to derive explicit quality control procedures from its decisions. We therefore focused on susceptibility-induced distortion, which is particularly evident near the rectum. Rectal cross-sectional area was measured in every patient using the Universal Segmenter (UniverSeg) model using a soft classification scheme trained on 40 manually segmented images.[24] Measurements were performed on the central slice of both T2 and ADC images. The Student t test was used to compare rectal cross-section between quality groups.

Statistical Analysis

The influence of individual acquisition parameters on image quality was assessed by a multi-pronged statistical approach using the binary classification scheme. The Kruskal-Wallis non-parametric rank test was used to compare median values across different quality categories. The Kolmogorov-Smirnov test assessed the overall distribution of parameters between passing and failing scans. Finally, piecewise logistic regression identified potential thresholds in acquisition parameters that could serve as quality control indicators. Dunn’s multiple comparisons test was used to reduce false positives to reduce false positives from multiple hypothesis testing. P-values were adjusted independently for each statistical test. Statistical calculations were performed using GraphPad Prism 10.1.2(324), except for piecewise regression which used the segmented R package (version 2-12).

Study Approval

The study was conducted under Institutional Review Board approval and in compliance with HIPAA regulations, with all participants providing written informed consent (ClinicalTrials.gov identifier: NCT03354416).

Results

Quality Analysis of the Study Population

Examples of high- and low-quality MRI scans for T2 images and ADC maps are shown in Figure 1A. There was a significant disparity in image quality between institutions, with 25% of ADC maps and 15% of T2 image sets from outside institutions rated as unsatisfactory, compared to 8% and 3% for the in-house images, respectively (Fig. 1C). To understand these differences, we analyzed representative high- and low-quality scans (Fig. 1A) and categorized the most common artifacts affecting image quality. Our analysis identified three primary types of artifacts in non-diagnostic images. Consistent with previous reports, [6, 25] susceptibility artifacts from rectal gas was found in the majority of the clinically non-diagnostic ADC images (Fig. S1A–C). While these artifacts were common in outside institution scans, they were rarely observed in in-house images. Low contrast affecting both DWI and ADC maps represented the second most common artifact type (Fig. S1D–F), while aliasing from improper field of view was observed less frequently (Fig. S1G–I). T2W image quality issues were less common overall (15% external vs 3% in-house) and primarily manifested as motion artifacts due to respiration (Fig. S1J).

Notably, subjective high image quality on T2W imaging did not guarantee similar quality in DWI/ADC maps. The quality of the preceding T2 scan was only moderately correlated with the quality of the ADC map during the same visit (r~0.4 for both sites) (Fig. 1C). This disconnect between T2 and ADC quality raises the question of whether these imaging failures reflect inherent patient characteristics or technical factors. To address this question, we examined the pattern of successful and failed scans across institutions:

Impact of Patient Anatomy on Image Quality:

The majority of patients were successfully imaged, with only 4 out of 235 T2 images and 9 out of 235 ADC images receiving a non-diagnostic quality score at both sites (Fig. 1B). This low number indicates that it was possible to image most patients successfully at least at one institution. There was a low but non-zero correlation (r~0.2) between paired scans of the same patient at different sites for both modalities (Fig. 1C), suggesting that while intrinsic patient characteristics played a role, they were not the primary determinant of overall MRI scan quality. However, the observed number of double failures for T2 images exceeded the expected number under the null hypothesis of independence (Fisher’s exact test, p<0.001, ADC p=0.07), suggesting that failures were not entirely independent events. Specifically, a review of representative slices from all cases where imaging was significantly compromised on both in-house and external sites suggested bladder distension, sometimes coincident with or secondary to benign prostatic hyperplasia, was particularly problematic for both T2 and ADC images (Fig. S2). Nevertheless, these cases represented exceptions rather than the norm, indicating that patient anatomy was not the main factor influencing MRI scan quality. This implies that with proper procedures, the vast majority of patients can be successfully scanned.

Impact of Technical Parameters on Image Quality

As our analysis revealed that patient anatomy was not the primary determinant of MRI scan quality, we conducted a more thorough investigation of technical acquisition parameters to determine their potential influence on image quality. Since the in-house dataset did not have significant image acquisition diversity, we focused our analysis on the multicenter dataset to better understand the role of technical parameters across diverse imaging environments. Our analysis of the multicenter external data confirmed previous findings that there was no association between PI-RADSv2 adherence and MRI quality for both T2 and ADC modalities (Fig. 3C and F). We then conducted a comprehensive statistical analysis of individual acquisition parameters from DICOM files in our multicenter dataset using three complementary approaches: the Kruskal-Wallis test to examine median differences across quality categories on a 3-point scale, the Kolmogorov-Smirnov test to compare parameter distributions between passing and failing scans (Figs 3A and B), and piecewise logistic regression to identify potential quality control thresholds. Despite this comprehensive analysis, no T2 acquisition parameter except a horizontal phase encoding direction emerged as statistically significant (logistic regression, p<0.001, AUC=0.6735), suggesting that T2 image quality is not largely determined by any single technical parameter in isolation.

Figure 3.

(A and B) Pairwise comparison of acquisition parameters between diagnostic (“Pass”) and non-diagnostic (“Fail”) quality images for T2-weighted images (A) and ADC maps (B). Except for the vertical resolution in ADC maps, no parameter was statistically significant in the multicenter dataset by Kolmogorov-Smirnov test with Holm-Sidak correction for multiple comparisons. (C) Association between PI-RADS v2 standards adherence and reader quality scores for T2-weighted images. (D) Pairwise comparison of the high B value of the ADC map. (E) ROC curve of the high B value (AUC=0.60) (F) Association between PI-RADS v2.1 standards adherence and reader quality scores for T2-weighted images. (G) Pairwise comparison of the number of B values used to make the ADC map. (H) ROC curve of the number of B values (AUC=0.52)

For ADC maps, results were similar, with only the vertical resolution (KS test, p=0.006, AUC= 0.6348) emerging as statistically significant. We specifically did not see a statistically significant difference with factors that had previously been shown to improve MRI quality including the number of B values (Fig.3D and E),[26] the FOV in either direction (Fig. 3A and B)[27, 28]. or the maximum B value (Fig.3G and H),[29–31] although the analysis of the latter was hampered by the low retrieval rate from the DICOM files (19%). This disparity likely stemmed from a combination of technical acquisition differences and patient preparation steps, making it challenging to isolate specific factors without considering potential interactions between imaging parameters.

Deep Learning-Based Multi-Stage Quality Assessment

Results of our quality assessment neural network (Fig. 2) are shown in Table 1 and Fig. 4. The first section shows the results of the model of each cross-validation fold against the validation set, while the bottom section is the best performing model in the test set against the validation set. In the multicenter validation set, ADC quality prediction from T2 image analysis achieved 65±13% sensitivity and 44±10% positive predictive value (PPV), with the best model reaching 83% sensitivity and 50%, with a notably high negative predictive value (NPV) of 90%. The ADC map assessment showed comparable performance (67% sensitivity, 62% PPV, 88% NPV), while the combined T2/ADC analysis yielded the best overall performance with 83% sensitivity, 50% PPV, and 93% NPV. Importantly, the T2 assessment step, which allows for immediate corrective actions, maintained high sensitivity (83%) and NPV (90%), enables early identification of potential quality issues. The model showed even stronger performance on the in-house dataset, achieving >90% accuracy across all modalities (T2: 92%, ADC: 94%, Combined: 94%), despite being trained exclusively on multicenter data. Additional experiments training ConvNext models on either multicenter or in house data alone showed no difference in accuracy when tested on data from the other site. The model showed similar performance whether trained on multicenter data and tested on in-house data or vice versa, with no significant difference in accuracy compared to within-site testing (training and testing on the same dataset, Table 2). This suggests robust generalization across different imaging environments. This robust generalization, combined with the high NPV across all conditions, suggests the model could effectively screen out poor-quality scans while minimizing unnecessary workflow interruptions.

Table 1:

Model performance metrics for quality prediction across different training and testing scenarios.

Multicenter
From …	Accuracy	NPV	PPV	Specificity	Sensitivity	AUC
T2	69 ± 8%	86 ± 6%	44 ± 10%	70 ± 13%	65 ± 13%	0.72 ± 0.05
ADC	69 ± 7%	88 ± 3%	45 ± 3%	69 ± 11%	72 ± 11%	0.79 ± 0.02
Combined	71± 6%	89 ± 3%	46± 7%	69 ± 8%	75 ± 6%	0.78 ± 0.02
Multicenter – Selected Model
From …	Accuracy	NPV	PPV	Specificity	Sensitivity	AUC
T2	60%	90%	37%	51%	83%	0.79
ADC	81%	88%	62%	86%	67%	0.78
Combined	74%	93%	50%	71%	83%	0.81
In House
From …	Accuracy	NPV	PPV	Specificity	Sensitivity	AUC
T2	90 ± 1%	98± 0%	46 ± 4%	92± 1%	75 ± 0%	0.90 ± 0.02
ADC	94 ± 3%	99 ± 1%	61 ± 14%	95 ± 3%	90 ± 14%	0.99 ± 0.01
Combined	94 ± 2%	98 ± 0%	64 ± 11%	96 ± 2%	75 ± 0%	0.96 ± 0.02
In House – Selected Model
From …	Accuracy	NPV	PPV	Specificity	Sensitivity	AUC
T2	92%	98%	50%	93%	75%	0.91
ADC	94%	98%	60%	95%	75%	0.98
Combined	94%	98%	60%	95%	75%	0.95

NPV: Negative Predictive Value, PPV: Positive Predictive Value, AUC: Area Under the ROC Curve

Figure 4.

Performance metrics of the quality prediction model based on an independent validation set. (A-C) ROC curves for the multicenter dataset showing T2, ADC, and combined T2/ADC model performance respectively. Blue dashed line indicates the cutoff. (D-F) Corresponding ROC curves for the in-house dataset demonstrating stronger performance across all modalities. (G) Distribution of model scores across quality categories in the multicenter dataset (*p<0.05, **p<0.01). (H) Sankey diagram illustrating the flow of cases through the quality assessment pipeline, showing the distribution of true/false positives and negatives at each stage

Table 2.

Cross-site validation of T2 image quality prediction. Performance metrics (mean ± standard deviation) for models trained and tested on different combinations of multicenter and in-house datasets. ‘From … to’ indicates training and test sets respectively.

T2
From …	Accuracy	NPV	PPV	Specificity	Sensitivity	AUC
Multicenter to Multicenter	73 ± 3%	87 ± 5%	48 ± 4%	75 ± 9%	67 ± 19%	0.75 ± 0.01
Multicenter to In House	68 ± 24%	96 ± 4%	19 ± 19%	68 ± 29%	65 ± 42%	0.87± 0.10
In House to In House	76 ± 24%	92 ± 3%	35 ± 38%	80 ± 26%	35 ± 14%	0.67 ± 0.19
In House to Multicenter	65% ± 6%	81 ± 3%	37 ± 6%	70 ± 11%	50 ± 14%	0.68 ± 0.03

NPV: Negative Predictive Value, PPV: Positive Predictive Value, AUC: Area Under the ROC Curve

Rectal Cross-Sectional Area as a Simple, Interpretable Quality Metric

Using the UniverSeg [24] to automatically measure rectal cross-sectional area in the central slice, we found significant correlations with image quality (Fig. 5) [32]. Rectal cross-sectional area on T2 images showed a statistically significant difference between the lowest and highest quality ADC maps (Kruskal-Wallis, p=0.006). This difference was even more pronounced when analyzing ADC maps retrospectively, with significant differences between the lowest and all other quality categories (Kruskal-Wallis, p=0.008 and p<0.0001) (Fig. 5). Logistic regression identified specific breakpoints at 618 mm² for T2 and 760 mm² for ADC maps (AUC 0.65 and 0.69 respectively), suggesting potential threshold values for quality control procedures.

Figure 5.

Automated rectal segmentation and its relationship to image quality. (A) Example T2 image with corresponding soft prediction from Universal Segmenter. (B) Example ADC map with corresponding segmentation prediction. (C) Violin plots showing the distribution of rectal cross-sectional area across T2 image quality scores, with significant difference between lowest and highest quality categories (*p=0.006). (D) Similar plots for ADC maps showing significant differences between lowest and other quality categories (**p=0.008, ****p<0.0001). (E,F) ROC curves for predicting image quality based on rectal cross-sectional area measurements, with optimal thresholds identified at 618 mm² for T2 and 760 mm² for ADC maps (AUC 0.65 and 0.69 respectively)

Discussion

Low quality mpMRI scans cause clinically significant prostate cancers to be missed or delayed in diagnosis. Importantly, low quality scans lead to skepticism over the value of mpMRI in general. Consistent with prior reports highlighting challenges in prostate MRI quality, our review across 62 external sites found substantial rates of non-diagnostic images: 15% for T2-weighted sequences and 25% for ADC maps. Addressing this issue requires understanding its origins. Our results suggest quality issues stem predominantly from site-specific technical and procedural issues rather than intrinsic patient characteristics, with notably lower in-house failure rates (3-8%) compared to external sites (15-25%). The rarity of complete imaging failure at both sites (4/235 T2, 9/235 ADC) indicates that most patients can be successfully imaged with proper technique. This underscores the significant potential for improving diagnostic quality through the refinement and consistent application of best practices across imaging centers.

Identifying effective practices is challenging, however, as our findings question the sufficiency of simply standardizing technical parameters. Our comprehensive analysis of technical acquisition parameters revealed surprisingly limited associations with image quality. Despite employing multiple statistical approaches, only horizontal phase encoding direction showed significance for T2 imaging (p<0.001, AUC=0.6735), while vertical resolution was the sole significant parameter for ADC maps (p=0.006, AUC=0.6348). Notably, adherence to PI-RADSv2 technical standards showed no significant correlation with image quality, challenging current quality control paradigms that rely primarily on protocol standardization. The absence of a significant correlation indicates that image quality is determined by complex interactions between multiple factors rather than individual acquisition parameters, necessitating a more sophisticated approach to quality assurance in prostate MRI. This does not imply that acquisition parameters are unimportant or that standardization efforts are generally ineffective.[33] Rather, it suggests that within the range of parameters typically employed across different sites, variations in individual settings may have less impact than other factors, particularly when considered in isolation. The lack of strong correlation could indicate that most centers already utilize broadly acceptable parameters, but achieving consistent diagnostic quality likely requires greater attention to broader procedural aspects, including patient preparation, which significantly influences common artifacts like susceptibility distortions.

To overcome the limitations of protocol-based quality control revealed by our analysis, we developed a deep learning quality assurance approach. Trained on data from 63 diverse sites to address known generalization challenges,[34–37] our multi-center neural network achieved accuracy comparable to single-site models for ADC quality prediction [10, 32, 38], with in-house performance matching the best single-site model by Cipollari et al (83% sensitivity, 100% specificity). [10] While our in-house results (75% sensitivity, 96% specificity) are similar, multi-center performance was lower (83% sensitivity, 71% specificity) (Table 1), suggesting that single-site models may rely on site-specific features that fail to generalize, producing high but non-transferable accuracy. [35, 39] This trade-off between site-specific accuracy and generalizability aligns with findings from Alis et al, whose multi-center study also showed reduced accuracy compared to single-site studies.[12] For efficient workflow integration where most scans are adequate, minimizing unnecessary interruptions is crucial. The model’s high NPV (93%) supports throughput by reliably identifying acceptable scans that donť need intervention. However, the 71% specificity indicates that interruptions for review of falsely flagged acceptable scans will still occur. Importantly, the high accuracy (94%) when applied to internal data suggests the model can be retrained for each center to address site-specific variations. This adaptability could potentially mitigate the generalization challenges observed in the multicenter dataset while maintaining the benefits of the deep learning approach.

A key advantage of our model lies in its ability to predict subsequent ADC quality directly from the initial T2 images, an approach not yet widely explored. Since T2 images are typically acquired first, this prediction offers an early warning before proceeding with the time-consuming DWI/ADC sequences. Such a warning creates an opportunity for timely interventions to potentially salvage scan quality (Fig 6). An early quality check after the T2 scan can trigger corrective actions. For instance, if significant rectal gas is indicated by the rectal cross-sectional area (e.g., Fig S1 A–C), attempts can be made to eliminate it (Fig 6) or use antispasmodics before acquiring DWI/ADC. Alternatively, a warning from the neural network might prompt adjustments to subsequent sequence parameters, like modifying the field of view if aliasing is suspected (e.g., Fig S1 G–I), or simply repositioning the patient. Beyond immediate corrective actions for individual patients, persistent quality warnings at an institution could signal underlying protocol deficiencies needing systematic review. Ultimately, leveraging this early T2-based assessment aims to optimize workflow efficiency and reduce the need for costly and inconvenient rescans.

Figure 6:

Proposed decision workflow for prostate MRI quality control. The process begins with prescan parameter verification, followed by T2-weighted image acquisition and quality assessment. Neural network inference predicts potential ADC quality from T2 images before proceeding to full DWI/ADC acquisition. Quality control decision points (diamonds) trigger specific interventions (red boxes) when quality thresholds are not met, with final advancement to radiologist review only after passing all quality checks.

Beyond the immediate benefits of T2-based prediction for optimizing workflow and resource allocation, a deeper understanding of factors influencing T2 and ADC quality can further refine our approach and identify specific targets for intervention. One prominent factor often associated with image quality is signal-to-noise ratio (SNR). Existing research shows mixed results on its true impact on both T2 and ADC images, raising questions about its effectiveness as a major indicator of quality. Belue et al found reconstruction by a generative adversarial network strongly increased SNR in T2 images but this increase in SNR rarely led to an increase in perceptual image quality when the reader was blinded to the reconstruction [40]. Similarly, Lee et al found no difference in perceptual image quality for T2, DWI, or ADC images after deep learning reconstruction, although high reader variability complicated the statistical interpretation of these results [41]. “Noise” as a subjective factor, however, ranks highly in subjective perceptions of MRI image quality [6]. This suggests the subjective perception of “noise” may result from other artifacts contributing to low contrast with different physical sources. Low contrast in the presence of high SNR (Fig. S1F) may be the result of the progressive misalignment of voxels due to geometric distortions at higher B values [42], potentially obscuring diagnostic features more critically than traditional noise measurements. In line with this observation, distortion from susceptibility differences from rectal gas, which are especially noticeable near the rectum (Fig. S1D–F), are a significant factor. Interestingly, we found that a simple measure of rectal area on the T2 image correlated with subsequent ADC image quality (Fig. 5C and E), offering a potential interpretable marker for this type of distortion. [32].

Conclusion

Some limitations of our study should be acknowledged. While our model demonstrates promising results in predicting ADC quality directly from T2 images, iťs crucial to acknowledge limitations in fully understanding the specific features it relies on for its predictions. Unlike established metrics like SNR, which have a clear physical basis, our deep learning model operates through complex internal processes, making it challenging to pinpoint the exact image characteristics driving its quality assessment. Second, our model rates the ADC map and not the underlying DWI images, which may be more critical for diagnosis [3]. Our quality interpretation was performed by only a single radiologist and interobserver variation analysis is therefore not possible. Finally, our model is based on perceptual quality and not improvement in accuracy in downstream tasks, which may have different requirements. Despite these limitations, our study demonstrates the strong potential of direct T2-based prediction for optimizing prostate MRI workflow and future research can address these limitations to further refine this approach [32]. In conclusion, we developed an interpretable AI model, which jointly utilizes T2 images and ADC maps to predict non-diagnostic ADC maps. Although our model demonstrated reasonable performance in a multi-center dataset, further evaluation across diverse clinical settings is needed to confirm its potential benefit in practice [43].

Data Availability Statement

The code used in this study is publicly available at https://github.com/jbrender/QCMRI. The imaging datasets used during the current study are available from the corresponding author upon reasonable request.

Data sharing statement:

The code used in this study is publicly available at https://github.com/jbrender/QCMRI. The imaging datasets used during the current study are available from the corresponding author upon reasonable request.