Impact of deep learning-based MRI reconstruction on image quality and apparent diffusion coefficient measurements in patients with intracranial tumors: a prospective intra-individual study

Introduction

Intracranial tumors are a diverse group of neoplasms that are linked to significant morbidity, mortality, and long-term neurological disabilities (1). Magnetic resonance imaging (MRI) is central to their diagnosis, treatment planning, and follow-up (2), and diffusion-weighted imaging (DWI) has become a key component of routine neuro-oncologic protocols (3). DWI, by probing the Brownian motion of water molecules, provides noninvasive information on tissue microstructure through the derived apparent diffusion coefficient (ADC). This technique has been widely used for tumor characterization and for assessing treatment-related changes (4).

Studies have shown that ADC values are inversely correlated with tumor cellularity in a wide range of brain and extracranial tumors (5). Lower ADC values typically indicate higher cell density and reduced extracellular space, while higher ADC values may suggest the presence of lower-grade tumor components, necrosis, or vasogenic edema. In neuro-oncology, ADC metrics are linked to tumor grade, treatment response, and patient outcomes in cases of diffuse gliomas and glioblastoma. Additionally, ADC-based thresholds have been proposed for risk stratification and survival prediction (6-8). In this context, reliable and stable ADC measurements across scanners, protocols, and reconstruction techniques are essential. Even small systematic biases can significantly impact lesion classification and compromise the comparability of longitudinal or multicenter data.

However, conventional brain DWI faces a well-known trade-off among spatial resolution, signal-to-noise ratio (SNR), geometric distortion, and acquisition time. While high b-value DWI can enhance lesion conspicuity, it often struggles with low SNR and susceptibility-related artifacts, especially near the skull base and air–bone interfaces (3). To maintain acceptable scan times in clinical practice, many protocols compromise spatial resolution or the number of signal averages. This compromise can degrade image quality and limit the robustness of quantitative diffusion metrics. Consequently, there is a clear need for reconstruction techniques that enhance DWI image quality without such compromises.

Recent advances in artificial intelligence have introduced deep learning–based MRI reconstruction algorithms in clinical settings (9,10). These methods, commonly implemented as convolutional neural networks within the reconstruction pipeline, are trained on extensive datasets of high- and lower-quality image pairs. Their purpose is to perform noise suppression and resolution enhancement either in the k-space or the image domain (11,12). Vendor-provided deep learning reconstruction (DLR) solutions have been evaluated in a variety of sequences and anatomical regions, and multiple studies (13,14) have consistently reported improved SNR and contrast-to-noise ratio, reduced blurring and truncation artifacts, and the potential for scan-time reduction while maintaining diagnostic performance. However, the technical details of commercial DLR algorithms are often not fully disclosed, which may limit mechanistic interpretation and external generalizability across vendors and platforms.

In diffusion imaging, DLR has demonstrated promise for applications in the prostate, liver, breast, and bladder. Specifically for prostate cancer, deep learning-based DWI reconstruction has been reported to significantly enhance image quality while maintaining the accuracy of ADC quantification at 3.0 T (15). In liver DWI, several studies (16,17) have shown that DLR improves SNR, enhances lesion conspicuity, and increases reader preference. Additionally, it has been associated with small yet significant changes in lesion ADC values when compared to conventional reconstruction. Similarly, deep learning-enhanced DWI for breast and bladder imaging offers superior subjective and objective image quality, and may also adjust ADC metrics or diagnostic thresholds in certain contexts (18,19). Together, these findings indicate that the effect of DLR on ADC values is not consistent and may vary depending on the organ being studied, the vendor’s implementation, and the acquisition protocol.

In the context of brain DWI, emerging evidence suggests that DLR can enhance image quality, particularly at higher b values, and may influence quantitative diffusion parameters. One study noted that DLR significantly improved the quality of brain DWI and altered several intravoxel incoherent motion (IVIM) indices, although ADC values were less significantly affected in that setting (20). Recent clinical studies on deep learning-accelerated and deep learning-reconstructed brain DWI have demonstrated improved perceived image quality and diagnostic confidence. However, these studies also indicate that ADC values derived from DLR images tend to be systematically lower in some settings compared to those from conventional reconstructions in various brain regions (21). Overall, these findings highlight the benefits of DLR while emphasizing the necessity for thorough quantitative validation in its application to brain diffusion imaging. At the same time, the potential advantages of DLR may be more apparent in technically demanding scenarios, such as high b-value imaging or multi-b-value acquisitions for advanced diffusion models, whereas its added value in routine single-b-value clinical DWI remains to be fully defined.

In neuro-oncology, ADC serves as a critical imaging biomarker, making it essential to address any reconstruction-induced bias, as that could significantly affect ADC-based thresholds, treatment response criteria, and longitudinal or multicenter studies of intracranial tumors. However, despite the swift clinical adoption of DLR, there is a scarcity of prospective intra-individual data regarding its quantitative impact in neuro-oncologic diffusion imaging, and most existing studies are limited in sample size, acquisition setting, or disease specificity, and have not focused specifically on intracranial tumors (20,21). Moreover, the potential clinical relevance of reconstruction-related ADC shifts in brain tumors remains insufficiently defined, particularly with respect to diagnostic interpretation and quantitative comparability.

This study aimed to prospectively assess the impact of a vendor-provided deep learning-based MRI reconstruction algorithm on DWI image quality—specifically background signal, noise, and lesion signal intensity (SI)—in patients with intracranial space-occupying lesions. We compared these results with those obtained from conventional reconstruction. Because the present study was performed under a conventional single-b-value DWI protocol (b=1,000 s/mm²), our primary objective was to evaluate the technical effect of DLR in a routine clinical setting rather than to establish its full utility in more advanced diffusion applications. By quantifying improvements in image quality and any systematic shifts in ADC, we sought to clarify the benefits and potential risks of using DLR for DWI in brain tumors. At the same time, this study was not designed to assess tumor grade classification, treatment response, or other downstream clinical endpoints. Additionally, our findings aim to support future standardization and calibration of quantitative neuro-oncologic imaging. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2768/rc).

Methods

Study population

This single-center, prospective, intra-individual controlled study was conducted at The First Affiliated Hospital of Chongqing Medical University. From January to May 2025, 58 consecutive adult patients diagnosed with intracranial space-occupying lesions and who underwent brain MRI were enrolled.

The inclusion criteria were: (I) age ≥18 years; (II) presence of an intracranial space-occupying lesion on imaging, with a maximum diameter >5 mm; and (III) availability of complete clinical and imaging data, along with the ability to cooperate with the MRI examination and provide informed consent. The exclusion criteria included: (I) contraindications to MRI, such as a cardiac pacemaker, suspected metallic foreign body, or critical metallic implants; (II) severe motion or susceptibility artifacts that resulted in non-diagnostic image quality; and (III) a history of prior brain surgery or radiotherapy that could significantly alter local diffusion characteristics.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All participants or their legal guardians provided written informed consent after being informed of the study’s purpose and potential risks. The study protocol was approved by the institutional ethics committee of The First Affiliated Hospital of Chongqing Medical University (No. 2024-265-01).

MRI acquisition protocol

All MRI examinations were conducted on the same 3.0 T scanner (Discovery MR750w, GE Healthcare, Waukesha, WI, USA) using a phased-array head coil. Patients were positioned supine with head-first entry. The standardized brain MRI protocol included axial T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), fluid-attenuated inversion recovery (FLAIR), and DWI.

DWI was acquired using a single-shot spin-echo echo-planar imaging sequence. The main acquisition parameters were as follows: repetition time (TR) of 4,000–5,000 ms, echo time (TE) of approximately 70 ms, slice thickness of 5.0 mm, interslice gap of 1.5 mm, acquisition matrix of 120×160, field of view (FOV) of 240 mm × 240 mm, and b values of 0 and 1,000 s/mm². Thus, the present study was performed under a conventional single-b-value DWI protocol (b=1,000 s/mm²), reflecting routine clinical practice in brain tumor imaging. The frequency and phase-encoding directions were consistent across all patients.

For each examination, the same raw k-space DWI data were reconstructed using two different methods: (I) a conventional reconstruction algorithm; and (II) a vendor-supplied deep learning-based reconstruction algorithm (IQ Engine, version 1.0, GE Healthcare), which employs a convolutional neural network for noise reduction and resolution enhancement. As this was a commercial vendor-provided reconstruction platform, detailed information regarding the internal network architecture, training dataset characteristics, number of parameters, and proprietary implementation was not fully available to the authors. Aside from the reconstruction algorithm, all other reconstruction parameters, including slice positions, matrix size, FOV, and filters, were identical between the two image sets. No high-b-value or multi-b-value diffusion acquisition was included in the present study, and reconstruction time was not systematically recorded.

Measurement of SI and background noise

All quantitative measurements were conducted using an AW 4.7 workstation (GE Healthcare). Due to the variability in lesion size and location among patients, the axial slice that best visualized the lesion and exhibited the largest cross-sectional area was selected for analysis. A neuroradiologist with over 10 years of experience in neuroimaging manually placed ROIs.

Three types of ROIs were defined on the DWI magnitude images:

• Lesion ROI: this ROI encompassed the solid portion of the lesion, intentionally excluding cystic, necrotic, or hemorrhagic areas, as well as cerebrospinal fluid spaces.
• Perilesional normal-appearing brain tissue ROI: this ROI was situated in morphologically homogeneous brain parenchyma adjacent to the lesion on the same side, while avoiding ventricles and visible vessels.
• Background ROI: this ROI was positioned in the extracranial background air region at the same slice level, approximately 2–4 cm from the scalp and near the skull, while steering clear of areas affected by fold-over artifacts or peripheral signal contamination.

For each ROI, the area was maintained between approximately 120 and 150 mm². To ensure spatial consistency between reconstruction methods, ROIs were initially drawn on one reconstruction method (either conventional or deep learning) and then copied and pasted to the corresponding slice of the other reconstruction. SI values were recorded for the lesion, perilesional brain tissue, and background ROIs. The standard deviation (SD) of the background ROI signal was measured and utilized as an estimate of background noise.

ADC measurement

ADC measurements were conducted on the AW 4.7 workstation utilizing the ADC calculation module within the READY View application. ADC maps were generated independently from both the conventionally reconstructed and deep learning-reconstructed DWI data.

For each patient, an ROI was placed within the solid component of the tumor on the ADC map, corresponding to the same slice used for DWI analysis. The placement of the ROI adhered to the same principles as those applied to the DWI magnitude images, carefully avoiding cystic, necrotic, and hemorrhagic regions, as well as any obvious perilesional vasogenic edema. The neuroradiologist measured the ADC three times at the same ROI location, with the mean of these three measurements serving as the final lesion ADC for each reconstruction method.

To reduce observer bias, the placement of the ROI and subsequent measurements were performed in a randomized order concerning the type of reconstruction. Additionally, the neuroradiologist was blinded to whether the images were conventionally or deep learning-reconstructed. In instances where there was uncertainty regarding the lesion boundaries, a second senior radiologist reviewed the ROI placement, and any discrepancies were resolved by consensus.

Statistical analysis

All statistical analyses were performed using SPSS Statistics version 26.0 (IBM Corp., Armonk, NY, USA). Continuous variables were presented as mean ± SD. The Shapiro-Wilk test was employed to evaluate the normality of data distributions. For normally distributed variables, paired t-tests were utilized to compare SI values of lesions and perilesional brain tissue, background noise (background SD), and lesion ADC values between conventional and DLRs. In contrast, Wilcoxon signed-rank tests were applied for variables that did not satisfy normality assumptions, serving as a sensitivity analysis.

All statistical tests were two-sided, with a P value <0.05 deemed statistically significant. When appropriate, relative percentage changes were calculated to quantify the impact of the reconstruction effect on each metric.

Results

Patient characteristics

A total of 63 patients with intracranial space-occupying lesions were assessed for eligibility. Five patients were excluded (3 due to postoperative status, 1 due to metal dental prostheses, and 1 due to prior radiotherapy), leaving 58 patients for the final analysis. No examinations were excluded due to non-diagnostic image quality (Figure 1). This cohort consisted of 34 women and 24 men, with ages ranging from 31 to 83 years and a mean age of 59.24±11.13 years. The lesions were classified based on imaging results and, where available, histopathology: neuroepithelial tumors (including astrocytoma, glioblastoma, and oligodendroglioma) were identified in 21 patients, brain metastases in 15 patients, meningiomas in 17 patients, and primary central nervous system lymphoma in 5 patients.

Figure 1 Flow diagram of patient enrollment, exclusions, and final study cohort.

Out of the 58 patients, 46 received histopathological confirmation following surgical resection or biopsy. Nine patients diagnosed with brain metastases through imaging, along with one patient diagnosed with a meningioma, declined surgery and were subsequently discharged; their diagnoses were based on typical imaging features and clinical context.

Comparison of DWI SI and background noise

Quantitative measurements of SI and background noise on DWI are summarized in Table 1 and illustrated in Figure 2.

Figure 2 Representative case of a pathologically confirmed meningioma with ROI placement on diffusion-weighted imaging. A 75-year-old female with a surgically and histopathologically confirmed meningioma. Axial DWI images reconstructed using conventional reconstruction (left) and deep learning-based reconstruction (right) are shown for the same slice position. Three circular ROIs were placed in an identical manner on both images: ROI 1 within the solid tumor component (lesion signal), ROI 2 in the adjacent perilesional brain parenchyma (perilesional tissue signal), and ROI 3 in the air region outside the head to estimate background signal/noise. ROI size and location were kept consistent between reconstructions to enable paired quantitative comparisons. DWI, diffusion-weighted imaging; ROI, region of interest.

For lesion SI, the mean SI for conventionally reconstructed images was 1,782.41±878.79, while the mean SI for deep learning-reconstructed images was slightly higher at 1,786.42±878.56. This difference was statistically significant (t=2.863, P=0.006), although the absolute change was small.

Regarding perilesional brain tissue, the mean SI was 1,446.99±243.22 with conventional reconstruction and 1,449.12±244.19 with DLR. The difference between these two methods was not statistically significant (t=−1.141, P=0.259).

In contrast, significant differences were noted in the background signal and noise. The mean background SI decreased from 34.95±4.99 with conventional reconstruction to 29.40±5.05 with DLR (t=13.948, P<0.001). Similarly, the mean background noise, represented as the SD of the background ROI, decreased from 20.61±4.17 to 14.63±3.50 (t=14.721, P<0.001), indicating an approximate 30% reduction in background noise with DLR.

Comparison of lesion ADC values

Lesion ADC values derived from conventional and deep learning-reconstructed DWI are shown in Figure 3.

Figure 3 Lesion ADC measurements on corresponding ADC maps reconstructed with conventional and deep learning methods. Representative axial ADC maps from the same patient and identical anatomical level are displayed, with conventional reconstruction (left) and deep learning-based reconstruction (right). A circular ROI (ROI 1) was placed within the solid tumor component in a matched position and size on both maps to obtain lesion ADC values. ADC, apparent diffusion coefficient; ROI, region of interest.

On conventionally reconstructed ADC maps, the mean lesion ADC was (1,324.41±617.97)×10⁻⁶ mm²/s. On deep learning-reconstructed ADC maps, the mean lesion ADC was significantly lower, at (1,247.10±567.37)×10⁻⁶ mm²/s. The two reconstruction methods showed a statistically significant difference (t=2.974, P=0.004). Specifically, the application of DLR resulted in an average decrease of approximately 6% in lesion ADC.

Discussion

In this prospective intra-individual study of patients with intracranial tumors, we found that deep learning-based MRI reconstruction significantly improved the quality of diffusion-weighted images while also introducing a modest but statistically significant systematic shift in lesion ADC values. Specifically, compared to conventional reconstruction, DLR reduced background signal and noise by approximately 30% and slightly increased lesion SI, thereby enhancing the conspicuity of lesions. However, the average lesion ADC values were about 6% lower in deep learning-reconstructed images than in conventionally reconstructed ones. These findings underscore the dual impact of DLR in neuro-oncologic diffusion imaging: it offers clear benefits for both qualitative and quantitative image quality metrics, while also introducing potential bias in quantitative ADC measurements. In the present study, DLR was evaluated only under a conventional single-b-value DWI protocol with b=1,000 s/mm², which reflects routine clinical practice in brain tumor imaging. Although this design allowed an intra-individual comparison between conventional reconstruction and DLR using the same raw data, it does not fully capture the potential value of DLR in more technically challenging diffusion settings.

Our results both align with and diverge from previous research on deep learning–based reconstruction in other organ systems. Ueda et al. (15) reported that DLR of prostate DWI at 3.0 T significantly enhanced image quality across various b values without substantially affecting ADC quantification. This finding suggests that, for their specific implementation and protocol, the main advantage of deep learning lies in noise reduction and artifact suppression rather than changes to diffusion metrics. Further studies in prostate MRI have supported the idea that DLR can enhance image quality or reduce acquisition time while largely maintaining the ADC values used in PI-RADS assessment (13).

In contrast, several studies of liver DWI have reported that DLR may subtly but significantly alter ADC values. Chen et al. (17) prospectively evaluated DLR in liver DWI and found that it improved SNR, contrast-to-noise ratio, and lesion conspicuity. However, they also noted small systematic changes in lesion ADC values when compared to conventional reconstruction. Afat et al. (16) found that deep learning–reconstructed liver DWI offered higher signal intensities on ADC maps and allowed for a reduction in acquisition time. However, the ADC values for parenchyma and lesions differed from those obtained using standard reconstruction. More recently, research on free-breathing liver DWI with DLR confirmed not only improved image quality and diagnostic performance but also highlighted measurable effects on ADC metrics (22).

For brain DWI, the literature remains relatively sparse but is rapidly growing. Hanamatsu et al. (20) investigated DLR in brain DWI using both phantom and in vivo data. They reported significant improvements in image quality, while noting that the ADC and pure diffusion coefficients were comparatively less affected in their experimental setting. Park et al. (21) demonstrated that deep learning-reconstructed brain DWI provides higher image quality. They also reported that the ADC derived from DLR was generally lower and exhibited less variability across multiple brain regions compared to the ADC obtained from conventional reconstruction. Phantom work (23) has demonstrated that DLR can influence ADC measurements, with the magnitude and direction of bias varying based on b values and slice thickness. Our findings expand this evidence to intracranial tumors, showing that in a clinical neuro-oncologic context, a vendor-provided DLR algorithm enhances brain DWI image quality while also introducing a modest, systematic downward shift in lesion ADC values.

Several mechanisms may account for the observed reduction in ADC values with DLR. First, diffusion-weighted MR images are inherently impacted by Rician noise (24-26), which can produce a noise floor that biases magnitude images upward, especially at high b values (27-29). Deep learning-based denoising integrated into the reconstruction pipeline can effectively reduce the noise floor in a non-linear and spatially varying fashion (30). Deep learning algorithms can suppress noise more effectively in high-b images than in low-b or b0 images. This capability may alter the relative signal attenuation across b values, leading to systematically lower ADC estimates in lesions characterized by strong diffusion weighting and limited SNR.

Second, many vendor-provided DLR algorithms are trained on pairs of low- and high-quality images to enhance perceptual image quality, edge sharpness, and texture appearance (13,31). In doing so, they may implicitly regularize spatial gradients and smooth subtle intensity variations within lesions, thereby reducing intra-lesional heterogeneity (32,33). Such smoothing can impact mono-exponential fits used to calculate ADC, particularly when only two b values are available. In our study, this effect may have contributed to the observed modest reduction in lesion ADC values. However, because we did not perform voxel-wise or histogram-based analysis, the influence of DLR on intralesional ADC distribution and heterogeneity could not be directly assessed.

Third, DLR may interact with other acceleration techniques, such as parallel imaging or compressed sensing, in ways that are not entirely clear to end users. Recent research on compressed sensing with deep learning-constrained reconstruction has shown that the choice of algorithms can either mitigate or introduce ADC bias, depending on the reconstruction model and regularization parameters (10,16,34,35). Because the present study used a commercial vendor-provided DLR implementation, detailed information regarding the internal network architecture, training dataset, and algorithmic integration with other reconstruction components was not fully available. In our cohort, the same acquisition protocol was applied for both conventional and DLR, indicating that the observed differences are primarily due to the reconstruction algorithm rather than to acquisition changes.

From a clinical perspective, the improvements in background noise, lesion signal, and overall image quality are beneficial and may enhance lesion detection, delineation, and reader confidence in patients with intracranial tumors. Enhanced conspicuity is especially valuable for lesions located near the skull base, small metastases, or infiltrative components of high-grade gliomas, where conventional DWI faces limitations due to susceptibility artifacts and low SNR (21,36). At the same time, the SI of perilesional tissue was not significantly altered by DLR in our study. One possible explanation is that the effect of DLR may depend on baseline tissue contrast and signal abnormality: lesions with more conspicuous diffusion-related signal changes may benefit more visibly from denoising and contrast enhancement, whereas perilesional tissue with less pronounced signal abnormality may show less measurable change. This interpretation, however, remains speculative and should be validated in future studies. However, the systematic reduction in lesion ADC values raises important considerations for neuro-oncologic applications that depend on quantitative diffusion metrics (37,38).

If DLR is introduced into routine practice without recalibrating these thresholds, there is a risk that lesions previously classified as “low ADC” or “high-grade” based on conventional reconstruction may be reclassified into different categories when imaged with DLR (39). This could affect clinical decision-making (40). However, the present study did not directly evaluate whether the observed ADC decrease altered tumor grade classification, diagnostic thresholds, treatment response assessment, prognostic stratification, or interobserver agreement in lesion characterization. Therefore, the clinical significance of the observed ADC bias should be interpreted cautiously. Longitudinal follow-up studies that utilize ADC to monitor treatment response may face confounding issues if the reconstruction techniques change over the course of follow-up (41,42).

Our results support the notion that implementation of DLR in neuro-oncologic diffusion imaging requires careful validation of ADC metrics. This should include consideration of reconstruction-specific reference ranges and, where feasible, harmonization strategies. Such strategies may involve cross-calibration with phantoms, statistical harmonization, or the use of conversion factors between reconstruction methods. In addition, phantom-based validation would be particularly valuable for determining whether the observed ADC shift is attributable to the reconstruction process itself rather than to physiological or lesion-related variability. Standardized diffusion phantom experiments may also improve the technical validation of DLR and facilitate cross-site and cross-vendor comparability of quantitative diffusion metrics.

Strengths and limitations

This study has several strengths. First, we employed a prospective design with intra-individual comparisons, ensuring that both conventional and DLRs were performed on the same raw DWI data for each patient. This approach minimizes confounding factors such as inter-scan variability, patient motion, and physiological changes, enabling a direct assessment of the effects of reconstruction. Second, we targeted a clinically relevant population of adults with intracranial space-occupying lesions, yielding data that can be directly applied to neuro-oncologic practice. Third, we assessed both qualitative surrogates of image quality (background signal, noise, lesion signal) and quantitative diffusion metrics (lesion ADC), allowing us to weigh the benefits and risks associated with DLR.

However, several limitations should be acknowledged. First, this was a single-center study with a relatively small sample size, and all examinations were performed on a single vendor platform using one specific DLR implementation. Therefore, the findings may not be generalizable to other vendors, field strengths, or reconstruction algorithms. Multi-center, cross-vendor validation will be needed to establish the reproducibility and generalizability of these results. Second, ROIs were manually drawn on lesions and perilesional regions. Although this approach reflects routine clinical practice and ROIs were copied between reconstruction methods to ensure spatial consistency, inter- and intra-observer reproducibility of the initial ROI delineation was not formally assessed. Third, we did not perform histology-specific subgroup analyses or voxel-wise/histogram-based analyses of ADC distributions. As different tumor types and grades may vary in cellularity, vascularity, necrosis, and diffusion characteristics, the pooled analysis may have masked heterogeneity in the effect of DLR across tumor subtypes. Fourth, the study included only a conventional single-b-value DWI protocol (b=1,000 s/mm²). At this b-value, standard DWI on modern 3.0 T scanners generally provides acceptable image quality, which may reduce the apparent incremental benefit of DLR. These scenarios were not evaluated in the present study. Accordingly, our findings should be interpreted as evidence of technical feasibility and quantitative stability in routine DWI rather than as a comprehensive demonstration of the full clinical utility of DLR. Finally, no diffusion phantom experiment was performed, and reconstruction time was not systematically recorded. As a result, we could not determine under standardized conditions whether the observed ADC shift was attributable to the reconstruction process itself rather than to physiological or lesion-related variability, nor could we assess the impact of DLR on clinical workflow efficiency. Future studies should incorporate standardized diffusion phantom experiments to support technical validation and improve cross-site and cross-vendor comparability of DLR-related quantitative effects.

Conclusions

In conclusion, under a conventional single-b-value DWI protocol (b=1,000 s/mm²), deep learning-based MRI reconstruction improved image quality in patients with intracranial tumors by reducing background noise and enhancing lesion conspicuity. However, this improvement was accompanied by a modest but systematic downward shift in lesion ADC values. Although the overall agreement of ADC measurements between reconstruction methods remained acceptable, this finding suggests that quantitative diffusion parameters may be influenced by the reconstruction process. Therefore, careful validation, standardization, and, where necessary, calibration of ADC metrics are essential when implementing DLR in clinical brain tumor diffusion imaging. Given that the present study did not assess downstream clinical endpoints and was limited to a routine single-b-value acquisition, the broader clinical utility of DLR, particularly in more technically demanding diffusion settings, remains to be established. Further validation in larger, clinically stratified, multicenter cohorts and across advanced diffusion techniques is needed before DLR can be more fully integrated into neuro-oncologic diffusion imaging workflows.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2768/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2768/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2768/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All participants or their legal guardians provided written informed consent after being informed of the study’s purpose and potential risks. The study protocol was approved by the institutional ethics committee of The First Affiliated Hospital of Chongqing Medical University (No. 2024-265-01).

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Armstrong TS. Intracranial tumors. Semin Oncol Nurs 2004;20:221-3. [Crossref] [PubMed]
2. Perkins A, Liu G. Primary Brain Tumors in Adults: Diagnosis and Treatment. Am Fam Physician 2016;93:211-7.
3. Romano A, Palizzi S, Romano A, Moltoni G, Di Napoli A, Maccioni F, Bozzao A. Diffusion Weighted Imaging in Neuro-Oncology: Diagnosis, Post-Treatment Changes, and Advanced Sequences-An Updated Review. Cancers (Basel) 2023;15:618. [Crossref] [PubMed]
4. Bulik M, Kazda T, Slampa P, Jancalek R. The Diagnostic Ability of Follow-Up Imaging Biomarkers after Treatment of Glioblastoma in the Temozolomide Era: Implications from Proton MR Spectroscopy and Apparent Diffusion Coefficient Mapping. Biomed Res Int 2015;2015:641023. [Crossref] [PubMed]
5. Surov A, Meyer HJ, Wienke A. Correlation between apparent diffusion coefficient (ADC) and cellularity is different in several tumors: a meta-analysis. Oncotarget 2017;8:59492-9. [Crossref] [PubMed]
6. Chen L, Liu M, Bao J, Xia Y, Zhang J, Zhang L, Huang X, Wang J. The correlation between apparent diffusion coefficient and tumor cellularity in patients: a meta-analysis. PLoS One 2013;8:e79008. [Crossref] [PubMed]
7. Chenevert TL, Malyarenko DI, Galbán CJ, Gomez-Hassan DM, Sundgren PC, Tsien CI, Ross BD. Comparison of Voxel-Wise and Histogram Analyses of Glioma ADC Maps for Prediction of Early Therapeutic Change. Tomography 2019;5:7-14. [Crossref] [PubMed]
8. Kurokawa R, Baba A, Kurokawa M, Capizzano A, Hassan O, Johnson T, Ota Y, Kim J, Hagiwara A, Moritani T, Srinivasan A. Pretreatment ADC Histogram Analysis as a Prognostic Imaging Biomarker for Patients with Recurrent Glioblastoma Treated with Bevacizumab: A Systematic Review and Meta-analysis. AJNR Am J Neuroradiol 2022;43:202-6. [Crossref] [PubMed]
9. Ottesen JA, Storas T, Vatnehol SAS, Løvland G, Vik-Mo EO, Schellhorn T, Skogen K, Larsson C, Bjørnerud A, Groote-Eindbaas IR, Caan MWA. Deep learning-based Intraoperative MRI reconstruction. Eur Radiol Exp 2025;9:29. [Crossref] [PubMed]
10. Dratsch T, Siedek F, Zäske C, Sonnabend K, Rauen P, Terzis R, Hahnfeldt R, Maintz D, Persigehl T, Bratke G, Iuga A. Reconstruction of shoulder MRI using deep learning and compressed sensing: a validation study on healthy volunteers. Eur Radiol Exp 2023;7:66. [Crossref] [PubMed]
11. Kim S, Park H, Park SH. A review of deep learning-based reconstruction methods for accelerated MRI using spatiotemporal and multi-contrast redundancies. Biomed Eng Lett 2024;14:1221-42. [Crossref] [PubMed]
12. Noordman CR, Yakar D, Bosma J, Simonis FFJ, Huisman H. Complexities of deep learning-based undersampled MR image reconstruction. Eur Radiol Exp 2023;7:58. [Crossref] [PubMed]
13. Kiryu S, Akai H, Yasaka K, Tajima T, Kunimatsu A, Yoshioka N, Akahane M, Abe O, Ohtomo K. Clinical Impact of Deep Learning Reconstruction in MRI. Radiographics 2023;43:e220133. [Crossref] [PubMed]
14. Rastogi A, Brugnara G, Foltyn-Dumitru M, Mahmutoglu MA, Preetha CJ, Kobler E, et al. Deep-learning-based reconstruction of undersampled MRI to reduce scan times: a multicentre, retrospective, cohort study. Lancet Oncol 2024;25:400-10. [Crossref] [PubMed]
15. Ueda T, Ohno Y, Yamamoto K, Murayama K, Ikedo M, Yui M, Hanamatsu S, Tanaka Y, Obama Y, Ikeda H, Toyama H. Deep Learning Reconstruction of Diffusion-weighted MRI Improves Image Quality for Prostatic Imaging. Radiology 2022;303:373-81. [Crossref] [PubMed]
16. Afat S, Herrmann J, Almansour H, Benkert T, Weiland E, Hölldobler T, Nikolaou K, Gassenmaier S. Acquisition time reduction of diffusion-weighted liver imaging using deep learning image reconstruction. Diagn Interv Imaging 2023;104:178-84. [Crossref] [PubMed]
17. Chen Q, Fang S, Yuchen Y, Li R, Deng R, Chen Y, Ma D, Lin H, Yan F. Clinical feasibility of deep learning reconstruction in liver diffusion-weighted imaging: Improvement of image quality and impact on apparent diffusion coefficient value. Eur J Radiol 2023;168:111149. [Crossref] [PubMed]
18. Cha SH, Han YE, Han NY, Kim MJ, Park BJ, Sim KC, Sung DJ, Yoo S, Lan P, Guidon A. Assessing Image Quality in Multiplexed Sensitivity-Encoding Diffusion-Weighted Imaging with Deep Learning-Based Reconstruction in Bladder MRI. Diagnostics (Basel) 2025;15:595. [Crossref] [PubMed]
19. Zhang X, Xu X, Wang Y, Zhang J, Hu M, Zhang J, Zhang L, Wang S, Li Y, Zhao X, Chen Y. Reduced field-of-view DWI based on deep learning reconstruction improving diagnostic accuracy of VI-RADS for evaluating muscle invasion. Insights Imaging 2024;15:139. [Crossref] [PubMed]
20. Hanamatsu S, Murayama K, Ohno Y, Yamamoto K, Yui M, Toyama H. Deep learning reconstruction for brain diffusion-weighted imaging: efficacy for image quality improvement, apparent diffusion coefficient assessment, and intravoxel incoherent motion evaluation in in vitro and in vivo studies. Diagn Interv Radiol 2023;29:664-73. [Crossref] [PubMed]
21. Park SI, Yim Y, Lee JB, Ahn HS. Deep learning reconstruction of diffusion-weighted brain MRI for evaluation of patients with acute neurologic symptoms. Sci Rep 2024;14:24761. [Crossref] [PubMed]
22. Yoon J, Lee Y, Yoon S, Sung J, Benkert T, Lee J, Park SH. Deep learning reconstruction of free-breathing, diffusion-weighted imaging of the liver: A comparison with conventional free-breathing acquisition. PLoS One 2025;20:e0320362. [Crossref] [PubMed]
23. Hayashi T, Kojima S, Ito T, Hayashi N, Kondo H, Yamamoto A, Oba H. Evaluation of deep learning reconstruction on diffusion-weighted imaging quality and apparent diffusion coefficient using an ice-water phantom. Radiol Phys Technol 2024;17:186-94. [Crossref] [PubMed]
24. Andersson JL. Maximum a posteriori estimation of diffusion tensor parameters using a Rician noise model: why, how and but. Neuroimage 2008;42:1340-56. [Crossref] [PubMed]
25. Gudbjartsson H, Patz S. The Rician distribution of noisy MRI data. Magn Reson Med 1995;34:910-4. [Crossref] [PubMed]
26. Aja-Fernandez S, Alberola-Lopez C, Westin CF. Noise and signal estimation in magnitude MRI and Rician distributed images: a LMMSE approach. IEEE Trans Image Process 2008;17:1383-98. [Crossref] [PubMed]
27. Lin YC, Huang HM. Denoising of multi b-value diffusion-weighted MR images using deep image prior. Phys Med Biol 2020;65:105003. [Crossref] [PubMed]
28. Wiest-Daesslé N, Prima S, Coupé P, Morrissey SP, Barillot C. Rician noise removal by non-Local Means filtering for low signal-to-noise ratio MRI: applications to DT-MRI. Med Image Comput Comput Assist Interv 2008;11:171-9. [Crossref] [PubMed]
29. Prah DE, Paulson ES, Nencka AS, Schmainda KM. A simple method for rectified noise floor suppression: Phase-corrected real data reconstruction with application to diffusion-weighted imaging. Magn Reson Med 2010;64:418-29. [Crossref] [PubMed]
30. Pfaff L, Darwish O, Wagner F, Thies M, Vysotskaya N, Hossbach J, Weiland E, Benkert T, Eichner C, Nickel D, Wuerfl T, Maier A. Enhancing diffusion-weighted prostate MRI through self-supervised denoising and evaluation. Sci Rep 2024;14:24292. [Crossref] [PubMed]
31. Wessling D, Gassenmaier S, Olthof SC, Benkert T, Weiland E, Afat S, Preibsch H. Novel deep-learning-based diffusion weighted imaging sequence in 1.5 T breast MRI. Eur J Radiol 2023;166:110948. [Crossref] [PubMed]
32. Sagawa H, Fushimi Y, Nakajima S, Fujimoto K, Miyake KK, Numamoto H, Koizumi K, Nambu M, Kataoka H, Nakamoto Y, Saga T. Deep Learning-based Noise Reduction for Fast Volume Diffusion Tensor Imaging: Assessing the Noise Reduction Effect and Reliability of Diffusion Metrics. Magn Reson Med Sci 2021;20:450-6. [Crossref] [PubMed]
33. Takayama Y, Sato K, Tanaka S, Murayama R, Goto N, Yoshimitsu K. Deep learning-based magnetic resonance imaging reconstruction for improving the image quality of reduced-field-of-view diffusion-weighted imaging of the pancreas. World J Radiol 2023;15:338-49. [Crossref] [PubMed]
34. Lemainque T, Yoneyama M, Morsch C, Iordanishvili E, Barabasch A, Schulze-Hagen M, Peeters JM, 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. [Crossref] [PubMed]
35. Bash S, Wang L, Airriess C, Zaharchuk G, Gong E, Shankaranarayanan A, Tanenbaum LN. Deep Learning Enables 60% Accelerated Volumetric Brain MRI While Preserving Quantitative Performance: A Prospective, Multicenter, Multireader Trial. AJNR Am J Neuroradiol 2021;42:2130-7. [Crossref] [PubMed]
36. Altmann S, Grauhan NF, Mercado MAA, Steinmetz S, Kronfeld A, Paul R, Benkert T, Uphaus T, Groppa S, Winter Y, Brockmann MA, Othman AE. Deep Learning Accelerated Brain Diffusion-Weighted MRI with Super Resolution Processing. Acad Radiol 2024;31:4171-82. [Crossref] [PubMed]
37. Tajima T, Akai H, Yasaka K, Kunimatsu A, Yamashita Y, Akahane M, Yoshioka N, Abe O, Ohtomo K, Kiryu S. Usefulness of deep learning-based noise reduction for 1.5 T MRI brain images. Clin Radiol 2023;78:e13-21. [Crossref] [PubMed]
38. Choi KS, Park C, Lee JY, Lee KH, Jeon YH, Hwang I, Yoo RE, Yun TJ, Lee MJ, Jung KH, Kang KM. Prospective Evaluation of Accelerated Brain MRI Using Deep Learning-Based Reconstruction: Simultaneous Application to 2D Spin-Echo and 3D Gradient-Echo Sequences. Korean J Radiol 2025;26:54-64. [Crossref] [PubMed]
39. Phuttharak W, Thammaroj J, Wara-Asawapati S, Panpeng K. Grading Gliomas Capability: Comparison between Visual Assessment and Apparent Diffusion Coefficient (ADC) Value Measurement on Diffusion-Weighted Imaging (DWI). Asian Pac J Cancer Prev 2020;21:385-90. [Crossref] [PubMed]
40. Padhani AR, Liu G, Koh DM, Chenevert TL, Thoeny HC, Takahara T, Dzik-Jurasz A, Ross BD, Van Cauteren M, Collins D, Hammoud DA, Rustin GJ, Taouli B, Choyke PL. Diffusion-weighted magnetic resonance imaging as a cancer biomarker: consensus and recommendations. Neoplasia 2009;11:102-25. [Crossref] [PubMed]
41. Kaufmann TJ, Smits M, Boxerman J, Huang R, Barboriak DP, Weller M, Chung C, Tsien C, Brown PD, Shankar L, Galanis E, Gerstner E, van den Bent MJ, Burns TC, Parney IF, Dunn G, Brastianos PK, Lin NU, Wen PY, Ellingson BM. Consensus recommendations for a standardized brain tumor imaging protocol for clinical trials in brain metastases. Neuro Oncol 2020;22:757-72. [Crossref] [PubMed]
42. Choi SJ, Kim KW, Ko Y, Cho YC, Jang JS, Ahn H, Kim DW, Kim MY. Whole Process of Standardization of Diffusion-Weighted Imaging: Phantom Validation and Clinical Application According to the QIBA Profile. Diagnostics (Basel) 2024;14:583. [Crossref] [PubMed]