AI‐Powered Gradient Echo Plural Contrast Imaging (AI‐GEPCI)—A Comprehensive Neurological Protocol From a Single MRI Scan

ABSTRACT

Background

MRI is essential for diagnosing and monitoring neurological diseases. Conventional protocols require multiple sequences to obtain complementary contrasts, increasing scan time, cost, and tolerability. Generating multiple contrasts from a single acquisition may streamline workflow while maintaining clinical utility.

Purpose

To train attention‐based convolutional neural networks (ACNNs) to generate clinical‐quality Fluid‐Attenuated‐Inversion‐Recovery (FLAIR), Magnetization‐Prepared‐Rapid‐Gradient‐Echo (MPRAGE), R2* maps, and derived contrasts from a single Gradient Echo Plural Contrast Imaging (GEPCI) acquisition.

Study Type

Retrospective.

Population

43 MRI scans from individuals with multiple sclerosis (25/18 F/M, 49 ± 11 years‐of‐age).

Field Strength/Sequence

3 T MRI, 3D GEPCI, MPRAGE, and FLAIR.

Assessment

Technical quality of AI‐generated contrasts was evaluated against directly acquired MRI using structural similarity index (SSIM). Clinical image quality was assessed by physicians. Lesion volumes and counts were obtained using automated segmentation.

Statistical Tests

One‐sample one‐sided Wilcoxon signed‐rank test was used to establish the clinical quality of images. Agreement between native‐ and AI‐derived lesion volume and lesion count measurements was assessed using intraclass correlation coefficients (ICC). Quantitative accuracy for R2* maps was evaluated using normalized root‐mean‐square error (NRMSE).

Results

AI‐generated FLAIR and MPRAGE achieved mean SSIM values of 0.923 ± 0.028 and 0.935 ± 0.022, respectively. Generated R2* maps achieved a mean SSIM of 0.996 ± 0.006 and NRMSE of 0.031 ± 0.020. Physicians‐assigned mean clinical quality ratings of 4.2 for GEPCI‐FLAIR and 4.5 for GEPCI‐MPRAGE exceeded the 4.0 clinical standard on a 1‐to‐5 scale. Lesion volume and count comparisons from automated segmentation showed strong agreement between AI‐generated and ground‐truth measurements: R² = 0.988 and R² = 0.933, ICC = 0.988 and ICC = 0.967, respectively.

Data Conclusion

AI‐GEPCI generated multiple clinically relevant MRI contrasts from a single GEPCI acquisition with high similarity to corresponding acquired images, supporting high‐quality, intrinsically co‐registered multi‐contrast brain evaluation.

Evidence Level

2.

Technical Efficacy

Stage 1.

Plain Language Summary

This study developed a new MRI method called AI‐GEPCI that can create many important brain images from a single scan. Normally, people with neurological diseases need several different MRI scans, which take a long time, are costly, and can be hard to tolerate. We trained artificial intelligence (AI) to turn one GEPCI scan into several clinical images (called FLAIR, MPRAGE, and R2*) that doctors routinely use. The AI‐generated images looked very similar to standard MRI images and were rated as high quality by physicians. This approach may allow faster, more comfortable, and more consistent MRI exams for brain diseases.

Abbreviations

ACNN

attention convolutional neural network

CVS

central vein sign

FLAIR

fluid attenuated inversion recovery

GEPCI

gradient echo plural contrast imaging

MPRAGE

magnetization‐prepared rapid gradient echo

MS

multiple sclerosis

NMAE

normalized mean absolute error

NRMSE

normalized root mean squared error

PRL

paramagnetic rim lesion

qGRE

quantitative gradient recalled echo

QSM

quantitative susceptibility mapping

SSIM

structural similarity index measure

SWI

susceptibility weighted imaging

1. Introduction

Machine learning (ML) methods have increasingly been applied to medical image reconstruction and analysis, providing data‐driven approaches for generating or enhancing MRI contrasts. A typical MRI examination consists of multiple pulse sequences, each providing a distinct image contrast sensitive to different tissue properties. Alternative approaches aim to reduce this burden either by using pulse sequences capable of generating multiple contrasts from a single acquisition, such as MR fingerprinting [1], and Gradient Echo Plural Contrast Imaging (GEPCI) [2], or by applying deep learning for automated MRI contrast synthesis and reconstruction [3, 4].

GEPCI is a post‐processing technique that extracts several image contrasts from a single Gradient Recalled Echo (GRE) MRI scan with mGRE pulse sequences [2] readily available on most commercial MRI scanners [2]. It was previously demonstrated that GEPCI reconstruction algorithms [2] allow the generation of images such as T1‐weighted, T2‐weighted Fluid‐Attenuated Inversion Recovery (FLAIR), and Susceptibility Weighted Imaging (SWI), along with quantitative R2* (1/T2*) maps, phase, and Quantitative Susceptibility Mapping (QSM) images [2]. An advanced version of GEPCI, quantitative Gradient Recalled Echo (qGRE) [5, 6], allows further separation of the R2* relaxation rate parameter into its subcomponents: R2t*, sensitive to tissue microstructure, and R2′, sensitive to vascular BOLD (Blood Oxygen Level Dependent) signal properties, both corrected for macroscopic field inhomogeneities [7] and physiological fluctuations [7, 8]. GEPCI and qGRE have proved useful in studying neurodegenerative diseases, such as Alzheimer's disease [9, 10], multiple sclerosis (MS) [8, 11, 12, 13, 14], psychiatric diseases [15], and traumatic brain injury [16].

While the original GEPCI pipeline employs a multi‐step process with model‐based fitting and investigator‐guided reconstructions, the AI‐GEPCI approach introduced herein leverages deep learning to automatically extract features from mGRE data and generate a suite of naturally coregistered clinical quality images and quantitative maps. This study aimed to demonstrate that the attention‐based convolutional neural networks (ACNNs) [17] can use the 10‐echo mGRE data from a single GEPCI acquisition as input to generate clinical‐quality FLAIR, Magnetization‐Prepared Rapid Gradient‐Echo (MPRAGE), and R2* images along with their derived contrasts, including T2*‐weighted images, FLAIR* [18], and introduced herein MP*.

As a specific example, we provide an AI‐GEPCI analysis for applications in multiple sclerosis (MS). Individuals with MS, a chronic inflammatory disease of the central nervous system characterized by demyelination, axonal injury, and neurodegeneration [19, 20, 21, 22], routinely undergo MRI for diagnosis and monitoring using sequences such as FLAIR, MPRAGE, SWI, and recently added FLAIR*, which together capture lesion visibility, structural detail, and overall lesion burden [23, 24]. This study demonstrates that GEPCI magnitude data can serve as the sole input for ACNN‐based synthesis of these clinically relevant contrasts within a single intrinsically co‐registered, multiparametric framework for MS lesion evaluation, including imaging features relevant to the central vein sign (CVS) and paramagnetic rim lesions (PRLs) [25, 26, 27, 28, 29, 30, 31, 32].

2. Materials and Methods

This retrospective cohort study was approved by the institutional review board (IRB), and informed consent was obtained from all participants. Eligible cases were identified retrospectively based on a documented clinical diagnosis of MS in the medical records and the availability of complete GEPCI, FLAIR, and MPRAGE acquisitions for analysis. The final cohort consisted of 43 MRI scans (25/18 F/M, 49 ± 11 years old). No additional inclusion or exclusion criteria were applied. The 43 available scans were partitioned by simple random assignment, without stratification, into training (n = 24), validation (n = 3), and testing (n = 16) subsets.

2.1. Imaging Protocol and Ground Truth Data Generation

MRI data were acquired with a Siemens‐3T Prisma scanner (Siemens Healthineers, Erlangen, Germany) equipped with a 32‐channel radiofrequency (RF) coil. MPRAGE and FLAIR were acquired with 1 × 1 × 1 mm³ voxel size and with matrices of 240 × 256 × 160 and 240 × 256 × 176, respectively. MPRAGE parameters were repetition time/echo time/inversion time (TR/TE/TI) = 2300/2.98/900 ms and flip angle (FA) = 9°, whereas FLAIR parameters were TR/TE/TI = 5000/393/1800 ms and FA = 120°. GEPCI was acquired with 1 × 1 × 1 mm³ voxel size, matrix of 192 × 256 × 160, TR = 50 ms, FA = 30°, 10 gradient echoes with first TE = 4 ms and echo spacing = 4 ms, pixel bandwidth = 450 Hz/pixel and Generalized Autocalibration Partial Parallel Acquisition (GRAPPA) [33] algorithm with acceleration factors of 2 in phase and slab directions.

After correction for physiological fluctuations [8] and GRAPPA reconstruction [33], the 32‐channel data were combined using a previously developed algorithm [2]. The quantitative R2* maps were produced by voxel‐by‐voxel fitting the biophysical model:

(1)

to the 10 GRE images. In Equation (1), TE is the gradient echo time, S ₀ is the signal amplitude, Δf is the local frequency shift, F _macro (TE) accounts for the adverse effects of macroscopic magnetic field inhomogeneities by means of a voxel spread function method [7].

The FLAIR and MPRAGE images were co‐registered to the GEPCI using the FMRIB Software Library (FSL) [34]. The magnitude of the complex GEPCI data was used and normalized by dividing all echoes by the mean value calculated across all the first echo images. The FLAIR and MPRAGE volumes were normalized to the mean values of their respective datasets.

2.2. Network Architecture and Training

For AI‐GEPCI, each subject's GEPCI magnitude images were provided to the network as a full 3D mGRE volume, implemented as a stack of transverse 2D slices with 10 echoes per slice. Three networks were trained to process data and generate three target 3D output volumes of FLAIR, MPRAGE, and quantitative R2* maps.

The ACNN networks (Figures 1, 2) consisted of down‐sampling encoder and up‐sampling decoder pathways, skip connections, and attention blocks [17]. The 3D volumes are loaded as sets of transverse 2D slices, and for each slice, the 10 echoes were provided together as input channels. The encoder extracted multi‐scale features, and the decoder up‐sampled and refined these feature maps using attention blocks and skip connections to combine information from corresponding encoder layers. The final convolutional layer generated a single reconstructed output 3D image volume. The attention blocks allowed the network to selectively emphasize echo‐dependent relaxation features relevant for contrast synthesis, particularly earlier echoes for MPRAGE‐like contrast and later echoes for FLAIR‐like contrast, while suppressing less informative signals.

FIGURE 1.

Network diagram for the attention U‐Net architectures utilized in this study [17]. The FLAIR, MPRAGE, and R2* map networks use the same architecture with different parameters. The networks take a 2D slice from a 3D GEPCI MRI volume with 10 gradient echoes as input. Each encoding block for the downsampling path performs two 3 × 3 convolutions, batch normalization, ReLu activation, and a 2 × 2 max pooling operation. The feature maps from each encoder block are passed forward via skip connections to the corresponding decoder block in the upsampling path. In the upsampling path, the decoder blocks utilize attention gates (detailed in Figure 2). The weighted skip connection from the attention gate and the two times upsampled output from the previous block are concatenated, have two 3 × 3 convolutions, batch normalization, and ReLu activation. The final decoder block includes an additional 1 × 1 convolution to reduce the 10 input channels to a single output image.

FIGURE 2.

Detailed attention gate architecture [17]. The attention gates receive the upsampled output from the previous decoder as the gating signal and the feature map from the corresponding encoder skip connection. Each map is processed individually through a 1 × 1 convolution, added elementwise, and activated by ReLu. The result is passed through another 1 × 1 convolution followed by a sigmoid function to produce a spatial attention map. This attention map is multiplied elementwise with the original skip connection, allowing only the most relevant features to pass to the decoder block via concatenation.

Two training configurations were evaluated: whole‐head inputs/labels and brain‐extracted inputs/labels. These were compared to determine whether restricting training to brain tissue improved synthesis performance versus preserving full anatomical context and avoiding a separate brain‐extraction preprocessing step. Unless otherwise specified, results reported in this work correspond to whole‐head training. Brain‐masked evaluation and figure display were used to focus technical and radiological assessment on intracranial image quality and to prevent non‐brain voxels from disproportionately influencing similarity metrics.

A post hoc ablation experiment compared the ACNN with an otherwise identical CNN lacking attention gates. Both models were trained with identical splits, loss functions, learning rates, and preprocessing. Performance was evaluated using Structural Similarity Index Measure (SSIM) [35] for the FLAIR, MPRAGE, and R2* outputs and normalized mean absolute error (NMAE) for the R2*.

Models were implemented in Python 3.10.13 using the PyTorch 2.0.0 and trained on a Windows 10 Enterprise workstation with an NVIDIA RTX A6000 GPU. Training was performed for 1000 epochs with batch size 2 using Adam optimization and required approximately 2 days per network. Evaluated loss functions were L2, L1, SSIM, and a 50/50 L2‐SSIM combination. Networks were trained with and without the FSL brain extraction tool (BET), and learning rates were selected based on training and validation losses. Validation losses were used for model selection during development, whereas all reported performance metrics correspond to the independent hold‐out test set. Learning rates were 10⁻⁵ for GEPCI‐to‐FLAIR and GEPCI‐to‐MPRAGE and 4 × 10⁻⁴ for GEPCI‐to‐R2*. After training, per‐subject inference required only a few seconds for structural contrast synthesis, with quantitative map generation in less than 1 s on the study hardware.

2.3. GEPCI‐Derived Images

Derived contrasts were generated from the ACNN‐produced battery of GEPCI images. T2* weighted images (T2*WI) (Equation 2), FLAIR*, and MP* images (Equation 3) can be computed for arbitrary echo time TE by using quantitative R2* maps:

$$T{2}^{*}\mathit{WI}=\exp \left(-R{2}^{*}·\mathit{TE}\right)$$ (2)

(3)

These TE‐dependent images were used to evaluate how echo time influences contrast behavior. Unless otherwise specified, examples of derived images in the Results section correspond to TE = 50 ms.

2.4. MS Lesion Evaluation

Lesion masks were generated using the automated Lesion Segmentation Tool (LST) for Statistical Parametric Mapping (SPM) (www.statistical‐modelling.de/lst.html), version 3.0.0. For most subjects, lesions were segmented using the lesion growth algorithm (LGA), which combines T1‐weighted tissue classification with FLAIR intensities to derive lesion probability maps [36]. For two subjects, the LGA did not yield adequate lesion segmentations, likely due to image quality limitations in the MPRAGE data; these cases were subsequently processed using the lesion prediction algorithm (LPA), which uses a logistic regression model trained on FLAIR data and spatial lesion priors [37].

2.5. Central Vein Sign (CVS) and Paramagnetic Rim Lesions (PRL)

CVS and PRL identifications were performed using qualitative visual evaluation based on established criteria for MS lesion characterization [30, 32]. Specifically, a lesion was considered to demonstrate a CVS if a small linear structure consistent with a central vein appeared in at least two FLAIR* imaging planes and appeared hypointense on T2*WI, FLAIR*, and MP*, and hyperintense on R2*. Correspondingly, a lesion was classified as a PRL if a peripheral rim was visible in at least two orthogonal planes and appeared hypointense on FLAIR*, T2*WI, and MP*, and hyperintense on R2*. CVS and PRLs were identified and counted on the AI‐GEPCI‐generated R2* maps and the derived T2*WI, FLAIR*, and MP* images.

2.6. Statistical Analysis and Image Evaluation

To assess the technical quality of the AI‐generated MRI images, image similarity was evaluated only within brain tissue by masking out non‐brain regions such as skull, scalp, and background. SSIM was computed between the generated FLAIR, MPRAGE, and R2* images (AI‐FLAIR, AI‐MPRAGE, and AI‐R2*) and their respective ground‐truth images. The accuracy of the quantitative AI‐R2* maps were additionally evaluated using the normalized root‐mean‐square error (NRMSE), calculated relative to the mean R2* value across the brain. For data analysis, 3D images were reconstructed in all three planes (axial, sagittal, and coronal). All metrics were computed slice‐by‐slice within each orientation and then aggregated to yield a single whole‐volume metric per subject. The results are reported as mean ± standard deviation for each test subject and across test subjects. While NRMSE was used for primary quantitative evaluation of R2* accuracy, NMAE was additionally employed in the ablation analysis to provide a complementary assessment of absolute error that is less sensitive to error squaring.

To evaluate lesion segmentation performance in the context of image‐quality assessment, lesion volumes and lesion counts were obtained using fully automated SPM‐based segmentation (LGA/LPA algorithms), and measurements from the AI‐GEPCI generated images were compared with those from the corresponding ground‐truth segmentations using scatter plots. Agreement between lesion volume and lesion count measurements obtained from the AI‐generated images and the corresponding ground‐truth segmentations was assessed using the coefficient of determination (R²) and intraclass correlation coefficients (ICC). The lesion statistics included in this study were provided solely for descriptive comparison to support image‐quality evaluation and were not intended as validation of a lesion‐segmentation method, which was beyond the scope of the present work.

To assess the radiological quality of the generated brain MRI images, the brain‐extracted AI‐FLAIR, AI‐MPRAGE, and corresponding ground‐truth FLAIR and MPRAGE volumes were reviewed by three physicians: one neuroradiologist with 10 years of experience (M.G.) and two MS neurologists with 40 and 20 years of experience, respectively (A.C. and G.W.). Each subject's GEPCI‐generated and corresponding ground‐truth image volumes were labeled as either A and B or B and A in pseudo‐randomized fashion and physicians independently reviewed each pair of blinded images according to overall quality, posterior fossa quality, and artifact quality on a standard subjective 1 to 5 Likert scale. For this study, scores were anchored based on physicians' experience as follows: 1 = nondiagnostic, 2 = poor/substantially below clinical quality, 3 = acceptable but below routine clinical quality, 4 = clinically routine diagnostic quality, and 5 = exceeding routine clinical quality. One‐sample one‐sided Wilcoxon signed‐rank testing against the clinical routine standard of 4.0 was performed for quality evaluation of AI‐generated and native images. Physicians were asked to identify the GEPCI‐generated images, and which (if either) was preferred. They also specified whether they could detect additional lesions on either the ground truth or GEPCI‐generated images.

3. Results

Performance of the AI‐GEPCI models for generating the AI‐FLAIR, AI‐MPRAGE images, and R2* maps from GEPCI data are summarized in Table 1. These data represent descriptive performance metrics across alternative loss functions and training configurations used for model selection. Among the evaluated loss functions, the L2 yielded the highest mean SSIM with lower variability for the AI‐FLAIR. For the AI‐MPRAGE, L1, L2, and the 50/50 L2‐SSIM combination produced similar SSIM values, whereas the SSIM‐only loss yielded lower mean SSIM values and greater variability across test subjects. For the quantitative AI‐R2* maps, only L1 and L2 loss functions were evaluated. BET L1 yielded a mean SSIM of 0.996 ± 0.006, whereas BET L2 yielded a mean SSIM of 0.980 ± 0.012 (Table 1).

TABLE 1.

Results obtained with different loss functions.

Loss function	FLAIR mean	FLAIR SD	MPRAGE mean	MPRAGE SD	R2* mean	R2* SD
L2	0.923	0.028	0.935	0.022	—	—
L1	0.914	0.039	0.941	0.024	—	—
SSIM	0.815	0.035	0.796	0.028	—	—
50% L2 + 50% SSIM	0.917	0.038	0.937	0.027	—	—
BET L2	0.888	0.058	0.899	0.061	0.980	0.012
BET L1	0.889	0.054	0.896	0.060	0.996	0.006

Note: Data show SSIM metric assessments (mean ± standard deviation across slices covering the brain volume) for networks trained using L2, L1, SSIM, 50% L2 & 50% SSIM loss functions. BET indicates training when only extracted brain was used as a ground truth. To ensure reliable metric calculations, the first and last few slices, often characterized by sparse data, were excluded from the analysis.

Abbreviations: BET, brain extraction tool; SSIM, structural similarity index measurement.

In the post hoc ablation comparison, the ACNN and non‐attention CNN produced similar descriptive results for GEPCI‐to‐FLAIR and GEPCI‐to‐MPRAGE synthesis. For GEPCI‐to‐R2*, the ACNN achieved a mean NMAE of 0.0127 ± 0.0077, whereas the non‐attention CNN achieved a mean NMAE of 0.0368 ± 0.0162. These descriptive results indicate lower reconstruction error for AI‐R2* generation with the ACNN architecture.

3.1. GEPCI To FLAIR and MPRAGE Results

The hold‐out test set consisted of 16 subjects processed through the GEPCI‐to‐FLAIR and GEPCI‐to‐MPRAGE networks. Across these subjects, the AI‐generated FLAIR images achieved a mean SSIM of 0.923 ± 0.028, and the AI‐generated MPRAGE images achieved a mean SSIM of 0.935 ± 0.022. The individual SSIM values are listed in Table 2. The relatively low SSIM variability across subjects indicates stable network performance. Representative ground‐truth and reconstructed images are shown in Figure 3a–c for FLAIR and Figure 3d–f for MPRAGE.

TABLE 2.

Subject‐wise quantitative agreement between AI‐GEPCI generated outputs and corresponding native reference images in the hold‐out test cohort.

Subject	FLAIR SSIM mean	SD	MPRAGE SSIM mean	SD	R2* SSIM mean	SD	R2* NRMSE mean	SD
1	0.886	0.022	0.935	0.036	0.9969	0.048	0.012	0.007
2	0.925	0.017	0.955	0.024	0.9992	0.0012	0.012	0.006
3	0.921	0.025	0.925	0.053	0.9962	0.010	0.019	0.016
4	0.943	0.029	0.943	0.042	0.9991	0.002	0.013	0.010
5	0.958	0.020	0.957	0.035	0.9963	0.013	0.018	0.022
6	0.961	0.022	0.961	0.031	0.9979	0.007	0.014	0.014
7	0.953	0.021	0.957	0.030	0.9993	0.0010	0.012	0.007
8	0.945	0.018	0.934	0.045	0.9752	0.059	0.040	0.058
9	0.945	0.018	0.953	0.032	0.9983	0.005	0.016	0.014
10	0.911	0.027	0.919	0.050	0.9984	0.003	0.015	0.012
11	0.938	0.025	0.949	0.016	0.9974	0.012	0.014	0.011
12	0.915	0.027	0.923	0.022	0.9964	0.014	0.017	0.012
13	0.893	0.019	0.951	0.017	0.9981	0.009	0.013	0.013
14	0.901	0.028	0.902	0.040	0.9952	0.011	0.091	0.042
15	0.871	0.028	0.911	0.032	0.9970	0.007	0.119	0.054
16	0.900	0.024	0.890	0.030	0.9932	0.015	0.078	0.029
Mean:	0.923	0.028	0.935	0.022	0.9960	0.006	0.031	0.020

Note: For each subject, SSIM is reported as mean ± standard deviation for FLAIR, MPRAGE, and R2* maps; R2* accuracy is additionally summarized by normalized root‐mean‐square error (NRMSE; mean ± SD). Subjects 1, 10, 13, and 15 had native clinical FLAIR and/or MPRAGE acquisitions with notable artifacts (examples shown in Figure 3); importantly, AI‐GEPCI reconstructions remained clinically interpretable despite these degradations, consistent with the artifact‐mitigation behavior highlighted in Figure 3.

FIGURE 3.

Representative examples of AI‐generated GEPCI outputs and corresponding ground‐truth images across three anatomical planes. For each contrast, FLAIR (a–c), MPRAGE (d–f), and quantitative R2* maps (g–i), the left image of each pair shows the ground‐truth reference and the right image shows the ACNN‐generated reconstruction. Across coronal, sagittal, and axial orientations.

A small subset of cases exhibited lower SSIM values and corresponded to ground‐truth images with visible artifacts (Figure S4). Because SSIM measures similarity to the available reference image, degradation in the ground‐truth images (not AI‐generated images) reduced measured agreement in these cases.

3.2. GEPCI To R2* Mapping Results

Across the 16 hold‐out test subjects, the AI‐generated R2* maps achieved a mean SSIM of 0.996 ± 0.006. Representative ground‐truth and reconstructed examples are provided in Figure 3g–i. As in other contrasts, SSIM was calculated only within the brain, and voxels outside the brain surface were excluded to avoid biasing the similarity measurements.

Quantitative accuracy of the AI‐GEPCI‐generated R2* maps was evaluated using NRMSE. The L1‐trained network produced a mean NRMSE of 0.031 ± 0.020, whereas the L2‐trained network yielded a mean NRMSE of 0.055 ± 0.025. These descriptive differences reflect the behavior of the respective loss method.

3.3. GEPCI Derived Images

Representative GEPCI‐derived images at TE = 50 ms are presented in Figure 4 along with corresponding AI‐FLAIR, AI‐MPRAGE, and AI‐R2* maps. Examples corresponding to different TEs are shown in Figures (S1–S3).

FIGURE 4.

Composite example of AI‐GEPCI generated and derived contrasts. Upper row: (a) AI‐FLAIR, (b) AI‐R2* map, and (c) AI‐MPRAGE images. Lower row: (d) FLAIR*, (e) T2*WI, and (f) MP* images for TE = 50 ms. Red arrows indicate representative periventricular lesions/CVS where applicable.

In these images, T2*WI demonstrated high signal in regions with lower R2* values, including CSF and some lesions. FLAIR* showed CSF suppression, increased lesion conspicuity, and clear gray matter–white matter contrast; MP* showed clear anatomical detail, well‐defined lesions, and suppressed CSF signal, suggesting that it may provide complementary information alongside FLAIR*. Representative examples of lesions demonstrating CVS and PRL imaging features across the generated and derived contrasts are shown in Figures 5 and 6.

FIGURE 5.

Central vein sign (CVS) example shown across axial, coronal, and sagittal planes in four AI‐GEPCI contrasts: (a) AI‐R2* map, (b) FLAIR*, (c) MP*, and (d) T2*WI images with TE = 50 ms. Red arrows indicate the same CVS lesion across planes and contrasts.

FIGURE 6.

Paramagnetic Rim Lesion (PRL) example in the sagittal plane across AI‐GEPCI generated and derived contrasts: (a) AI‐FLAIR, (b) R2* map, (c) AI‐MPRAGE, (d) FLAIR* with TE = 50 ms, (e) T2*WI with TE = 50 ms, and (f) MP* contrasts, using TE = 100 ms. Red arrows indicate the PRL lesion.

3.4. Radiological Assessment

The results of the radiological evaluations of AI‐generated and native MPRAGE and FLAIR images are summarized in Table 3. Mean overall ratings (measured on a 1‐to‐5 Likert scale) were 4.2 for AI‐FLAIR and 4.5 for AI‐MPRAGE, both exceeding 4.0, anchored as the clinically routine imaging quality standard. AI‐FLAIR and AI‐MPRAGE ratings for the posterior fossa were 4.2 and 4.3, respectively, while artifact ratings were 4.5 for both. One‐sample one‐sided Wilcoxon signed‐rank testing against the clinical routine standard of 4.0 showed that all native image categories and nearly all AI‐generated image categories were rated significantly above the clinical routine threshold (Table S1), with the exception of AI‐FLAIR posterior fossa ratings, which were higher than the clinical reference value of 4.0 but did not reach statistical significance (p = 0.05664).

TABLE 3.

Expert physician assessment of image quality of AI‐FLAIR, AI‐MPRAGE, and native FLAIR and MPRAGE.

Clinical assessment rating type	FLAIR mean	SD	AI‐FLAIR mean	SD	MPRAGE mean	SD	AI‐MPRAGE mean	SD
Overall rating	4.71	0.46	4.21	0.58	4.90	0.31	4.54	0.50
Posterior fossa rating	4.62	0.53	4.15	0.50	4.83	0.38	4.34	0.52
Artifact rating	4.73	0.49	4.52	0.62	4.79	0.50	4.50	0.68

Note: Ratings ranged from 1 through 5, with 5 being the best and 4 corresponding to the clinically routine imaging quality standard.

For the AI‐identification task, reader performance differed by contrast (Table S2). For FLAIR, correct identification rates were 87.5%, 25.0%, and 87.5% for the three readers (66.67% overall), indicating mixed reader performance. For MPRAGE, correct identification rates were 93.8%, 100%, and 62.5% (85.42% overall), indicating that the AI‐generated image was more consistently identified. While most lesions were equivalently seen on both scans, readers more often reported additional small or faint lesions on native than on AI‐generated FLAIR images, with mean counts of 7.3 versus 4.3 of 16 cases across readers. For MPRAGE, such differences were less frequent, with mean counts of 3.3 for native and 1.0 for AI‐generated images.

3.5. GEPCI Network Application in MS Lesion Visualization

Representative examples of periventricular, juxtacortical, CVS, and PRL lesion features across the AI‐GEPCI generated and derived contrasts are shown in Figures 4, 5, 6. Figure 5 presents an example of a CVS lesion visible in all three imaging planes across the AI‐R2* map and the derived MP*, FLAIR*, and T2*WI images. Figure 6 illustrates a representative PRL. The PRL appears as a bright partial ring on the AI‐R2* map (highlighted by a red arrow) and as a dark partial ring on the FLAIR*, MP*, and T2*WI images.

Quantitative comparison of lesion volumes and lesion counts between the AI‐GEPCI pipeline and ground‐truth segmentations showed strong agreement, with R² values of 0.98 and 0.933 and ICC values of 0.9876 and 0.9673 for lesion volume and lesion count, respectively (Figure 7).

FIGURE 7.

Comparison of AI‐GEPCI lesion segmentations to ground truth across lesion volumes and lesion counts. (a) Scatter plot of lesion volume comparison displays the AI‐GEPCI (AI) lesion volumes versus the ground truth (GT) lesion volumes. (b) Scatter plot of lesion count comparison from AI and GT segmentations. Each point represents a single test subject. The red lines represent the regression lines. The regressions equations and coefficients of determination (R²) are shown. The high R² values indicate strong correlations.

A subset of subjects exhibited reduced ground‐truth image quality in FLAIR, MPRAGE, or both due to motion‐ or acquisition‐related artifacts (example in Figure S4). These cases coincided with lower SSIM values in Table 2 and affected segmentation‐based comparison metrics.

4. Discussion

In this study, ACNNs were developed to generate a suite of naturally co‐registered, clinical‐quality images from a single mGRE GEPCI acquisition. These include FLAIR, MPRAGE, R2*, T2*WI, FLAIR*, and MP*. The generated images demonstrated excellent agreement with the corresponding clinical images, with SSIM indicating high correspondence for FLAIR, MPRAGE, and R2* maps and also excellent agreement between AI‐R2* maps and the biophysical model derived reference R2*, Equation (1). In blind assessments, three clinicians judged the GEPCI‐derived FLAIR and MPRAGE images to be comparable to or better than clinically routine FLAIR and MPRAGE images and only slightly lower in quality than the high‐quality native FLAIR and MPRAGE images used in this study. Moreover, the GEPCI‐derived images performed comparably with regards to detecting lesions while showing utility for CVS and PRL identification, not available from FLAIR and MPRAGE. In more complex cases, such as those with severe atrophy, extensive lesion burden, or numerous small closely spaced lesions, small differences in lesion count reflected lesion confluence or boundary separation rather than missed lesion tissue reflected in better agreement for the total lesion volume. These findings highlight the potential clinical feasibility of this approach. Importantly, this work leverages the inherent efficiency and versatility of GEPCI, reducing acquisition time and ensuring perfect natural co‐registration across multiple contrasts, an advantage over existing methods that rely on multiple separate sequences. The ability to generate both structural and quantitative maps from a single acquisition has strong implications for neurological disease imaging, particularly in MS, where accurate lesion visualization, tissue characterization, and quantitative monitoring are critical. Though further optimization of the GEPCI‐derived FLAIR and MPRAGE images would be helpful, these results suggest that they might already be clinically suitable. Both L1‐ and L2‐trained networks showed low variability across subjects, although the L1‐trained model exhibited lower NRMSE values for R2*, suggesting a possible advantage for this loss function in quantitative mapping.

The ablation analysis supports the use of attention mechanisms for quantitative R2* reconstruction. While both architectures produced visually comparable results, the ACNN achieved markedly lower normalized mean absolute error compared with the non‐attention CNN. This suggests that attention gates are particularly beneficial for quantitative GEPCI‐derived maps, where accurate modeling of subtle echo‐dependent signal variations is critical.

GEPCI is a single‐acquisition framework capable of generating multiple intrinsically co‐registered contrasts from the same underlying mGRE dataset. In this study, we focused on GEPCI outputs derived from magnitude data, which are widely available on commercial MRI scanners, and demonstrated that these data could serve as the sole input for AI‐based generation of clinically relevant structural and quantitative contrasts. These findings are important because they demonstrate the ability of AI‐GEPCI to expand the range of possible GEPCI outputs (such as MP*) and accelerate access to clinically useful GEPCI‐derived information within a practical workflow.

The developed ACNNs translated 10‐echo GEPCI mGRE data into a set of naturally co‐registered contrasts, including FLAIR, MPRAGE, R2* maps, FLAIR*, MP*, and T2*WI. A practical advantage of this approach is that all direct and derived outputs are generated from a single GEPCI acquisition, avoiding reliance on multiple separately acquired input sequences and preserving intrinsic co‐registration across contrasts. In the context of this study, this property is particularly relevant for MS, where lesion‐sensitive structural and quantitative contrasts may benefit from direct spatial correspondence across modalities.

Among the generated contrasts, MP* represents a novel addition that draws conceptual inspiration from FLAIR*, which combines FLAIR and T2*WI imaging to enhance lesion and vein conspicuity in MS. While FLAIR* improves the detection of lesions and the central vein sign through voxel‐wise combination of FLAIR and T2*WI data, MP* builds upon structural information from MPRAGE by multiplying it with T2*WI echoes across multiple echo times. This synthesis captures both fine anatomical detail and echo‐dependent pathological features, mimicking how FLAIR* highlights lesions and veins, while retaining the spatial resolution and contrast of MPRAGE. Notably, MP* contrast can be modulated by selecting the echo time: using an early echo (e.g., TE≈1 ms) emphasizes MPRAGE‐like structural detail, while later echoes (e.g., TE > 30 ms) enhance lesion conspicuity comparable to FLAIR*, offering a flexible diagnostic perspective from the same data. In the representative PRL example shown in Figure 6, MP* provided particularly clear conspicuity of the rim lesion relative to several other generated contrasts, suggesting that it may offer complementary value for depicting susceptibility‐related lesion features. Since MP*, FLAIR*, and T2*WI are all derived from the same single GEPCI acquisition and are inherently co‐registered, they collectively provide a rich multi‐contrast framework for lesion characterization.

Since 2018, various networks have been developed with the aim of synthesizing various image contrasts. Some models aim at data imputation [38, 39], and direct translations from various contrast combinations [4, 40, 41, 42, 43, 44, 45]. These approaches typically require multiple scan acquisitions as network inputs. In contrast, MRI methods like MR fingerprinting [1, 46], GEPCI [2], and Synthetic MR [47], allow generating images faster, with varying contrasts from a single sequence, improving efficiency and consistency of contrast generation.

From a practical perspective, improved lesion visualization is crucial in MS, where monitoring disease progression, particularly in terms of new or active lesions, is essential for treatment planning. Additionally, GEPCI‐derived images offer the potential for improved tissue characterization and the ability to track microstructural changes over time, offering insights into disease progression that could lead to more personalized therapeutic strategies [19, 20, 21, 22]. As previously demonstrated, the plurality of GEPCI‐generated images also makes this approach valuable for a wide range of clinical and research applications beyond MS, including neurodegeneration in Alzheimer's disease [9, 10], psychiatric diseases [15], and traumatic brain injury [16]. The ability to generate structural and quantitative maps from a single, rapid acquisition may support advancing understanding of these diseases and could inform future approaches to patient care.

5. Limitations

The retrospective dataset was obtained from a single center, using a single vendor, a single scanner, a single field strength and a relatively small number of MRI scans. Since radiological evaluation indicated that the quality of GEPCI‐derived FLAIR and MPRAGE images could benefit from further optimization, validation across larger and more diverse datasets will be required to assess broader applicability. The complex phase data were not explored but will be incorporated in future model development to expand GEPCI‐derived outputs.

6. Conclusion

This study demonstrated the successful application of ACNNs to generate a suite of clinically relevant, co‐registered MR images from a single GEPCI acquisition. The AI‐GEPCI approach provides a comprehensive set of naturally co‐registered imaging contrasts critical for neurological disease diagnosis and monitoring, particularly in conditions like MS. The promising performance in lesion detection, including CVS and PRL, suggests that GEPCI‐derived images could be a valuable addition to clinical practice, offering a cost‐effective and efficient alternative to current multi‐scan protocols. Since mGRE pulse sequences are readily available on practically all commercial MRI scanners, the AI‐GEPCI approach has great potential for broad implementation in clinical practice.

Funding

This work was supported by the National Institutes of Health (NIH) and National Institute on Aging (NIA) grants RF1AG077658, RF1AG082030, R01AG077658, and R01AG082030.

Supporting information

TABLE S1: One‐sample one‐sided Wilcoxon signed‐rank analysis of qualitative image ratings compared with the clinical routine imaging standard of 4.0. Tests were performed separately for native and AI‐generated FLAIR and MPRAGE images across overall quality, posterior fossa quality, and artifact quality ratings. Mean rating, median rating, and p value are reported for each category.

Table S2: Reader accuracy for identification of AI‐generated versus native image volumes in the blinded radiological assessment. Correct identifications are reported for each reader and overall as counts and percentages for FLAIR and MPRAGE.

Figure S1: T2*WI example montage. These T2*WI examples are calculated from the generated R2* map results using a TE of 35 ms, 50 ms, 70 ms, and 100 ms for the left column and progressing to the right.

Figure S2: FLAIR* example montage. These FLAIR* examples are calculated from the R2* results using the TE values of 35 ms, 50 ms, 70 ms, and 100 ms. TE increases from the left column to the right column.

Figure S3: MP* example montage. These MP* examples are calculated from the R2* results using the TE values of 35 ms, 50 ms, 70 ms, and 100 ms. TE increases from the left column to the right column.

Figure S4: Representative lower‐agreement examples showing degraded native reference image quality. (a) Native MPRAGE and (b) AI‐generated MPRAGE at matched slice location. This is an example when the native reference images exhibited reduced overall image quality characteristics, contributing to lower similarity metrics despite preservation of overall anatomical structure in the AI‐generated images.

Data Availability Statement

Data generated or analyzed during the study and ML training programs are available from the corresponding author by request.