Deep Learning-Powered Super Resolution Reconstruction Improves 2D T2-Weighted Turbo Spin Echo MRI of the Hippocampus

Abstract

Purpose: To assess the performance of 2D T2-weighted (w) Turbo Spin Echo (TSE) MRI reconstructed with a deep learning (DL)-powered super resolution reconstruction (SRR) algorithm combining compressed sensing (CS) denoising and resolution upscaling for high-resolution hippocampal imaging in patients with (epileptic) seizures and suspected hippocampal pathology. Methods: A 2D T2w TSE coronal hippocampal sequence with compressed sense (CS) factor 1 (scan time 270 s) and a CS-accelerated sequence with a CS factor of 3 (scan time 103 s) were acquired in 28 patients. Reconstructions using the SRR algorithm (CS 1-SSR-s and CS 3-SSR-s) were additionally obtained in real time. Two readers graded the images twice, based on several metrics (image quality; artifacts; visualization of anatomical details of the internal hippocampal architecture (HIA); visibility of dentate gyrus/pes hippocampi/fornix/mammillary bodies; delineation of gray and white matter). Results: Inter-readout agreement was almost perfect (Krippendorff’s alpha coefficient = 0.933). Compared to the CS 1 sequence, the CS 3 sequence significantly underperformed in all 11 metrics (p < 0.001-p = 0.04), while the CS 1-SRR-s sequence outperformed in terms of overall image quality and visualization of the left HIA and right pes hippocampi (p < 0.001-p < 0.04) but underperformed in terms of presence of artifacts (p < 0.01). Lastly, relative to the CS 1 sequence, the CS 3-SRR-s sequence was graded worse in terms of presence of artifacts (p < 0.003) but with improved visualization of the right pes hippocampi (p = 0.02). Conclusion: DL-powered SSR demonstrates its capacity to enhance imaging performance by introducing flexibility in T2w hippocampal imaging; it either improves image quality for non-accelerated imaging or preserves acceptable quality in accelerated imaging, with the additional benefit of a reduced scan time.

1. Introduction

The hippocampus is considered a strong epileptogenic area within the human brain [1]. High-resolution MRI of the hippocampus can identify and potentially exclude morphological abnormalities of the hippocampus such as hippocampal sclerosis or limbic encephalitis and has thus emerged as a valuable diagnostic modality in patients with epileptic seizures [1,2].

The MRI anatomy of the hippocampus is complex and intricate [1,3,4,5,6]. T2-weighted (T2w) imaging plays an important role in comprehensively visualizing the fine anatomical structures of the hippocampus and in potentially detecting the subtle changes indicative of pathology.

To this extent, high-resolution, thin-slice T2w TSE images are routinely required and acquired in clinical practice. However, such sequences are time-consuming, easily requiring more than five minutes of scan time, presenting economic implications due to limited scanner availability but also placing a burden on patients, who must remain motionless in an uncomfortable environment—making scan time reduction a clinical and logistical priority. With recent advances in MRI technology, including compressed sensing/compressed SENSE (CS) [7], image acquisition can be shortened considerably without compromising image quality or diagnostic accuracy. Specifically, these MRI acceleration technologies can be leveraged to either reduce scan time or improve spatial resolution without prolonging scan time [8].

Recently, artificial intelligence (AI)-powered deep learning (DL) MRI reconstruction technologies have been proposed to further enhance sequences acquired by CS techniques. Specifically, deep neural networks have shown promise for the efficient reconstruction of images from highly undersampled k-space data, as is the case for CS-accelerated MRI [9,10,11]. Exemplarily, an approach using a convolutional neural network, called adaptive-CS-Net [12,13,14], has shown potential to improve imaging across a variety of anatomical regions relative to standard compressed sensing [10,15,16,17,18]. By combining this approach with a second denoising DL neural network (Precise-Image-Net), removal of ringing artifacts and upscaling of image resolution can be achieved [12,19,20,21,22,23]. In the following, this combination is referred to as the super resolution reconstruction (SRR) algorithm. While this algorithm has successfully improved imaging of the prostate [12,24], breast [25], knee [23,26], and heart [27], its application to the hippocampus remains unexplored.

Thus, based on the considerations outlined above and the potential of the SRR algorithm, we sought to present and compare the clinical performance of a high-resolution coronal 2D T2w TSE sequence, reconstructed either with standard CS or SRR, for hippocampal imaging in patients with suspected epileptic seizures. Specifically, we sought to determine whether SRR offers advantages over conventional CS reconstruction by improving image quality at a fixed scan time or by enabling shorter scan times without compromising image quality—thereby enhancing the overall balance between acquisition time and diagnostic imaging performance.

2. Materials and Methods

2.1. Study Subjects

In this institutional review board (IRB)-approved intra-individual comparison study, we recruited 28 consecutive patients referred for routine hippocampal MRI as part of a comprehensive seizure evaluation for suspected epileptic seizures. Patients included were those presenting with first-time focal seizures (motor seizures, absences) or grand mal seizures for epilepsy diagnostic workup. The study was conducted at a single tertiary medical center between January and May 2024. For all patients, the dedicated epilepsy protocol with specific consideration of hippocampal imaging was acquired. Exclusion criteria included (1.) age < 18 years; (2.) incomplete data acquisition. Finally, 28 patients (mean age, 57 years; age range, 19 to 83 years; 16 male, 12 female) were included in the study. No patient had to be excluded. The final diagnoses were normal hippocampus (23 patients), mesial sclerosis (2 patients), and atrophy of the hippocampus due to neurodegenerative Alzheimer’s disease (3 patients). All patients provided informed consent.

2.2. Magnetic Resonance Imaging

Imaging was performed on a 3T MR Ingenia Elition scanner (Philips Healthcare, Best, the Netherlands) in supine position using a 32-channel transmit/receive head coil.

A detailed overview of the sequence parameters is provided in

Table 1. Sequence parameters.

	2D T2w TSE CS 1/CS 3
Acquisition	coronal
Drive pulse	no
Repetition time TR in ms	3668
Echo time TE in ms	90
Flip angle	90°
Echo train length ETL	15
Number of echoes	1
FOV	230 × 184 × 88
Slice thickness/gap	2 mm/0.2 mm
Acquired voxel size	0.6 mm
Reconstructed voxel size	0.4 mm
Number of slices	40
Acceleration	CS factor 1/CS factor 3
Number of signal averages NSA	2
Acquisition time	270 s (4 min 39 s)/103 s (1 min 43 s)

. Besides the conventional non-accelerated high-resolution coronal 2D T2w TSE sequence with a 2 mm slice thickness (reference standard; CS factor 1), an additional, identical sequence accelerated with a CS factor of 3 was acquired. For the second sequence, a CS factor of 3 was selected, as preliminary volunteer scans indicated that higher CS factors led to noticeable degradation in image quality. Lastly, for the purpose of the study, both sequences were additionally reconstructed in real time using a DL-powered SRR algorithm with the denoising level set to strong.

The reconstruction algorithm is vendor-provided (Philips Healthcare, Best, the Netherlands) and is implemented online. This newly developed DL framework combines and integrates compressed sensing with two distinct convolutional neural networks. The first network (the so-called Adaptive-CS-Net) for sparsity-constrained reconstruction with non-uniform random subsampling has already been established in several CS-AI studies [10,16,17] and mimics the iterative shrinkage–thresholding algorithm approach first described by Pezzotti et al. [13,14].

A second network addresses the removal of ringing artifacts and upscaling of image resolution [12,19]. A detailed description of the algorithm and its implementation can be found elsewhere [12,23,24,25,26,27].

2.3. Qualitative Image Analysis

MR images were evaluated by two readers (a board certified neuroradiologist with 28 years of experience and a neuroradiology resident with four years of experience in imaging research) in consensus during two readout sessions separated by at least 4 weeks in a blinded and randomized manner.

A set of hippocampal T2w TSE images of the same patient containing the 2D T2w TSE CS 1 (named CS 1), the 2D T2w TSE CS 3 (named CS 3), the 2D T2w TSE CS 1 with DL-based image reconstruction SRR strong s (named CS 1-SRR-s), and the 2D T2w TSE CS 3 with DL-based image reconstruction SRR strong s (named CS 3-SRR-s) were each assigned a number from 1 to 4. This set of images was prepared by a coworker of the study in a four-column format adjusted in random order and saved as anonymized (i.e., blinded) snapshots in our PACS system, allowing for scrolling through the sequences. These PACS snapshots were then evaluated by the two blinded readers. Readers evaluated the sets of images for all patients.

The readers subjectively assessed image quality using semi-quantitative Likert scales partially adapted from previous studies [3]. The following metrics were graded, i.e., overall image quality; presence of artifacts; and finally, delineation, conspicuity, and clarity of the following anatomical structures: hippocampal internal architecture, left and right; dentate gyrus, left and right; pes hippocampi, left and right; mammillary bodies; fornix; gray and white matter of the temporal gyri.

Artifacts were classified on a 4-point Likert scale: 1, severe; 2, moderate; 3, mild; 4, absent. Overall image quality was assessed on a 5-point Likert scale: 1, no diagnostic value; 2, very limited diagnostic value; 3, acceptable diagnostic value; 4, good diagnostic value; 5, optimal diagnostic value. The delineation, conspicuity, and clarity of the anatomical structures were classified on a 4-point Likert scale: 1, not visible; 2, poor clarity (blurred); 3, good clarity; 4, excellent clarity, with excellent delineation and well defined borders.

2.4. Statistical Analysis

To assess reliability among the readouts and due to the ordinal structure of the data, inter-readout agreement was quantified using Krippendorff’s alpha coefficients. Agreement was interpreted as follows: 0.00–0.20 = poor, 0.21–0.40 = fair, 0.41–0.60 = moderate, 0.61–0.80 = substantial, and 0.81–1.00 = almost perfect agreement.

Differences in image quality ratings across reconstruction methods were assessed using the non-parametric Friedman test for repeated measures. In the case of significant results, post hoc pairwise comparisons were performed using Wilcoxon signed-rank tests. Resulting p-values were corrected for multiple testing using the Benjamini–Hochberg procedure to control the false discovery rate. All statistical tests were two-tailed, and p-values < 0.05 were considered statistically significant. Data are presented as median (interquartile range) for ordinal variables and mean ± standard deviation (SD) for continuous variables. All analyses were performed using R statistical software (version 4.0.2; R Foundation for Statistical Computing, Vienna, Austria; https://www.R-project.org/, accessed on 1 June 2024).

3. Results

3.1. Scan Time

The 2D T2w TSE sequences CS 1/CS 1-SRR-s exhibited a scan time of 279 s, while the CS 3/CS 3-SRR-s sequences displayed a scan time of 103 s. Thus, the CS 3/CS 3-SRR-s sequences exhibited a scan time reduction of 63% as compared to that for the 2D T2w TSE sequences sCS 1/CS 1-SRR-s.

3.2. Image Analysis

Image examples are shown in Figure 1 and Figure 2. A detailed overview of the scores is provided in

Table 2. Overview of scores from subjective image analysis. S1 and S2 refer to readout sessions 1 and 2, respectively. Scores are displayed as median (interquartile range); mean ± standard deviation. Superscript stars refer to the level of significance (* = p < 0.05, ** = p < 0.01, *** = p < 0.001) for pairwise statistical comparisons against CS 1 as the reference standard.

	CS 1	CS 3	CS 1-SRR-s	CS 3-SRR-s
Overall Image Quality	S1: 4 (4,5); 4.2 ± 0.6 S2: 4 (4,4); 4.1 ± 0.6	S1: 3.5 (3,4); 3.5 ± 0.6 ** S2: 3 (3,4); 3.5 ± 0.6 **	S1: 5 (4,5); 4.4 ± 0.7 * S2: 5 (4,5); 4.5 ± 0.7 **	S1: 4 (4,4.25); 4.1 ± 0.6 S2: 4 (4,4.25); 4.1 ± 0.6
Presence of Artifacts	S1: 4 (3,4); 3.4 ± 0.7 S2: 4 (3,4); 3.4 ± 0.7	S1: 3 (2,3); 2.8 ± 0.6 ** S2: 3 (3,3); 2.8 ± 0.6 **	S1: 3 (3,4); 3.1 ± 0.8 ** S2: 3 (3,4); 3.1 ± 0.8 **	S1: 3 (3,3); 2.9 ± 0.7 ** S2: 3 (3,3); 2.9 ± 0.7 **
GWM in Temporal Gyri	S1: 4 (4,4); 3.9 ± 0.4 S2: 4 (4,4); 3.8 ± 0.4	S1: 3 (3,3); 3.1 ± 0.6 *** S2: 3 (3,3); 3.1 ± 0.5 ***	S1: 4 (4,4); 3.9 ± 0.4 S2: 4 (4,4); 3.9 ± 0.3	S1: 4 (3,4); 3.7 ± 0.5 S2: 4 (3,4); 3.6 ± 0.6
Fornix	S1: 4 (4,4); 3.8 ± 0.5 S2: 4 (4,4); 3.8 ± 0.5	S1: 3 (3,4); 3.2 ± 0.6 *** S2: 3 (3,2); 3.1 ± 0.5 ***	S1: 4 (4,4); 3.9 ± 0.4 S2: 4 (4,4); 3.9 ± 0.3	S1: 4 (4,4); 3.8 ± 0.5 S2: 4 (4,4); 3.8 ± 0.4
Hippocampal Internal Architecture Left	S1: 4 (3,4); 3.5 ± 0.7 S2: 4 (3,4); 3.5 ± 0.7	S1: 3 (2,3.25); 2.9 ± 0.8 ** S2: 3 (2,3.25); 2.9 ± 0.8 **	S1: 4 (4,4); 3.8 ± 0.6 ** S2: 4 (4,4); 3.8 ± 0.6 **	S1: 4 (3,4); 3.3 ± 0.8 S2: 4 (3,4); 3.3 ± 0.8
Hippocampal Internal Architecture Right	S1: 4 (3,4); 3.6 ± 0.8 S2: 4 (3,4); 3.6 ± 0.8	S1: 3 (3,4); 3 ± 0.9 ** S2: 3 (3,4); 3 ± 0.9 **	S1: 4 (4,4); 3.8 ± 0.8 S2: 4 (4,4); 3.8 ± 0.8	S1: 4 (3,4); 3.3 ± 0.9 S2: 4 (3,4); 3.4 ± 0.9
Mammillary Bodies	S1: 4 (4,4); 3.9 ± 0.5 S2: 4 (4,4); 3.8 ± 0.5	S1: 4 (3,4); 3.5 ± 0.6 * S2: 3 (3,4); 3.4 ± 0.7 **	S1: 4 (4,4); 4 ± 0.2 S1: 4 (4,4); 4 ± 0.2	S1: 4 (4,4); 3.9 ± 0.4 S1: 4 (4,4); 3.9 ± 0.3
Pes Hippocampi Left	S1: 4 (4,4); 3.9 ± 0.4 S2: 4 (4,4); 3.8 ± 0.4	S1: 4 (3,4); 3.5 ± 0.6 * S2: 3 (3,4); 3.4 ± 0.6 **	S1: 4 (4,4); 4 ± 0 S2: 4 (4,4); 4 ± 0	S1: 4 (4,4); 4 ± 0.2 S2: 4 (4,4); 4 ± 0.2
Pes Hippocampi Right	S1: 4 (4,4); 3.8 ± 0.4 S2: 4 (4,4); 3.8 ± 0.4	S1: 3.5 (3,4); 3.5 ± 0.5 * S2: 3 (3,4); 3.4 ± 0.5 **	S1: 4 (4,4); 4 ± 0 * S2: 4 (4,4); 4 ± 0 *	S1: 4 (4,4); 4 ± 0 * S2: 4 (4,4); 4 ± 0 *
CA3/4 Dentate Gyrus Left	S1: 4 (4,4); 3.8 ± 0.5 S2: 4 (4,4); 3.8 ± 0.5	S1: 3 (3,4); 3.4 ± 0.7 ** S2: 3 (3,4); 3.4 ± 0.7 **	S1: 4 (4,4); 3.9 ± 0.4 S2: 4 (4,4); 3.9 ± 0.4	S1: 4 (3,4); 3.6 ± 0.7 S2: 4 (3.75,4); 3.6 ± 0.7
CA3/4 Dentate Gyrus Right	S1: 4 (4,4); 3.8 ± 0.6 S2: 4 (4,4); 3.8 ± 0.6	S1: 4 (3,4); 3.5 ± 0.7 S2: 4 (3,4); 3.5 ± 0.7 *	S1: 4 (4,4); 3.8 ± 0.6 S2: 4 (4,4); 3.8 ± 0.6	S1: 4 (4,4); 3.8 ± 0.6 S2: 4 (4,4); 3.8 ± 0.6

and Figure 3.

Overall inter-readout agreement across all sequences and metrics was almost perfect, with alpha = 0.933. On a per-sequence basis, inter-readout agreement was also substantial and ranged from alpha = 0.883 (CS 1 sequence) to alpha = 0.959 (CS 1-SRR-s sequence). As for the scores of subjective image quality metrics, there were significant differences between the sequences.

In terms of overall image quality, the CS 3 sequence was deemed significantly worse (p < 0.004), while the CS 1-SRR-s sequence was deemed significantly better (p < 0.04) than the CS 1 sequence. The CS 3-SRR-s sequence was rated comparable to the CS 1 sequence (p > 0.6).

In terms of the presence of artifacts, the CS 1-SRR-s, CS 3, and CS 3-SRR-s sequences all performed worse than did the CS 1 sequence (p < 0.003).

For delineation of GWM in the temporal gyri, the CS 3 sequence performed worse than the CS 1 sequence (p < 0.001), while the CS 1-SRR-s and CS 3-SRR-s sequences performed comparably well (p > 0.06).

For the visibility of the fornix, the CS 3 sequence performed worse than the CS 1 sequence (p < 0.001) while the CS 1-SRR-s and CS 3-SRR-s sequences performed comparably well (p > 0.1).

For the delineation of the hippocampal internal architecture on the left side, the CS 3 sequence performed worse (p < 0.004), while the CS 1-SRR-s sequence performed better (p < 0.005) than the CS 1 sequence. The CS 3-SRR-s sequence was rated comparable to the CS 1 sequence (p > 0.3).

For the delineation of the hippocampal internal architecture on the right side, the CS 3 sequence was graded worse than the CS 1 sequence (p < 0.004), while the CS 1-SRR-s and CS 3-SRR-s sequences were deemed comparable (p > 0.06) to the CS 1 sequence.

Concerning the visualization of the mammillary bodies, the CS 3 sequence was rated worse than the CS 1 sequence (p < 0.01), while the CS 1-SRR-s and CS 3-SRR-s sequences were deemed comparable (p > 0.3).

As for the visualization of the pes hippocampi on the left side, the CS 3 sequence was rated worse than the CS 1 sequence (p < 0.01), while the CS 1-SRR-s and CS 3-SRR-s sequences were deemed comparable (p > 0.1).

As for the visualization of the pes hippocampi on the right side, the CS 3 sequence was rated worse than the CS 1 sequence (p < 0.02), while the CS 1-SRR-s and CS 3-SRR-s sequences were deemed significantly better (p = 0.02).

Concerning the visibility of the CA3/4 dentate gyrus on the left side, the CS 3 sequence was rated worse than the CS 1 sequence (p ≤ 0.008), while the CS 1-SRR-s and CS 3-SRR-s sequences were deemed comparable (p > 0.2).

Concerning the visibility of the CA3/4 dentate gyrus on the right side, the CS 3 sequence was rated comparable (p = 0.14) for the first readout but worse (p = 0.04) for the second readout compared to the CS 1 sequence. The CS 1-SRR-s and CS 3-SRR-s sequences were deemed comparable (p > 0.6) to the CS 1 sequence for both readouts.

4. Discussion

In this study we assessed the potential of a DL-powered SRR algorithm for improved T2w imaging of the hippocampus in patients with (epileptic) seizures and suspected hippocampal pathology. We showed that DL-powered SSR can enhance imaging performance by introducing flexibility in T2w hippocampal imaging. Specifically, it either improves image quality for non-accelerated imaging or preserves acceptable quality in accelerated imaging yet at the benefit of a reduced scan time.

The hippocampus exhibits an exceptionally intricate and delicate anatomy, requiring high-quality visualization to effectively evaluate any potential pathological changes. Specifically, the anatomy of the hippocampus on imaging can be described as follows: The Ammon horn of the hippocampus (cornu ammonis), with the T2w hyperintense pyramidal cell layers CA1 to CA4, surrounds the T2w hyperintense central dentate gyrus. The cornu ammonis is covered by white matter, seen as the T2w hypointense alveus that ends medially as the T2w hypointense fimbria. The Ammon horn of the hippocampus is in continuity with the subiculum and the subsequent entorhinal cortex of the parahippocampal gyrus. Between the dentate gyrus and the cornu ammonis CA1–CA4 and the subiculum, the T2w hypointense dark band of white matter of the stratum radiatum, lacunosum, and moleculare (SRLM) is located, thus forming a typical spiral appearance [3,4,6,28]. Thus, the hippocampal internal architecture (HIA) is based on this laminar structure of the body of the hippocampus, with a spiral appearance in the coronal plane due to the adjacent layers of gray and white matter in the Ammon’s horn [1,3]. This hypointense band is only 1mm thick in coronal images, which corroborates the necessity of acquiring high resolution and high quality T2w TSE images for clear visualization [3]. Importantly, changes in the HIA are key indicators of hippocampal sclerosis, making its clear visualization crucial for epilepsy patients.

Advancing hippocampal imaging for clinical purposes has not yet gained widespread attention. While literature on high-resolution imaging of the hippocampus for research purposes exists, these studies often focus on different imaging sequences, higher field strengths (e.g., 7T MRI), and significantly different requirements [4,29,30]. In research, protocols and scan times do not need to be as short as those required in clinical practice, meaning acceleration and efficiency are less critical. However, in a clinical setting, protocols must balance speed, high resolution, and sufficient signal-to-noise ratio to enable accurate diagnosis and the detection of potential pathologies. Despite its importance, there is a noticeable lack of studies specifically addressing the optimization of T2w imaging of the hippocampus at 3T for clinical applications.

Regarding the use of AI, a recent study by Suh et al. is of interest. The authors analyzed the diagnostic performance and image quality of three MRI protocols for temporal lobe epilepsy (TLE) in 117 patients. They found that 1.5 mm slice thickness MRI, with deep learning-based reconstruction (1.5-mm MRI + DLR), significantly improved pooled sensitivity (91.2% vs. 72.1%) and depiction of hippocampal abnormalities compared to routine 3 mm MRI, although with lower specificity (76.5% vs. 89.2%). Furthermore, 1.5 mm MRI + DLR provided higher accuracy (91.2% vs. 73.1%), as well as superior image quality, SNR, and CNR than did 1.5 mm MRI without DLR. The study concluded that 1.5 mm MRI + DLR enhances TLE diagnosis, particularly for hippocampal evaluation, due to improved lesion conspicuity and image quality [31].

Largely corroborating these findings, we also present an elegant deep learning-based solution that enhances T2w hippocampal imaging at 3T, achieving improved image quality without extending scan time or compromising other technical MRI parameters. However, in contrast to the approach by Suh et al. that employed deep learning to reduce noise from images, thus boosting image quality [31], our SRR algorithm goes further by aiming to reduce ringing artifacts and upscale image resolution.

Specifically, in our study, we demonstrated that for standard CS reconstruction, a three-fold acceleration of the CS sequence (from CS1 to CS3) resulted in a noticeable decline in image quality. At very high resolutions, such as those achieved here with voxel sizes of 0.6 mm, even slight imperfections become immediately apparent, further degrading image quality. Maintaining a sufficient signal within each voxel is essential for constructing reliable images. As a result, even minor accelerations pose significant challenges, as they inevitably reduce the signal-to-noise ratio, impacting the overall image quality.

However, incorporating the SRR algorithm into the CS3 sequence restored performance to a level comparable to that of the non-accelerated CS1 sequence, while achieving a considerable reduction in scan time of 63%. Alternatively, imaging performance could be further enhanced by applying the SRR algorithm to the CS1 sequence. Visualization of the hippocampal internal architecture and pes hippocampi was improved. Although a slight increase in artifacts was noted—likely attributable to the reconstruction process—the CS1-SRR-s sequence still provided overall superior image quality. Thus, integrating the SRR algorithm into clinical practice offers dual benefits, i.e., reduced scan time and enhanced image quality. While this study did not directly assess diagnostic outcomes, future validation should examine SRR’s impact on diagnostic performance, particularly since improved image quality demonstrably enhances both diagnostic reliability in hippocampal MRI, as also exemplarily demonstrated by Suh et al., as well as segmentation accuracy.

Hippocampal segmentation, increasingly used not only for research but also for clinical applications (e.g., Alzheimer’s diagnostics; epilepsy evaluation; and potentially, radiotherapy planning), is essential for quantifying biomarkers like atrophy [32]. Beyond increasing the sensitivity for detecting subtle lesions in epilepsy, advanced imaging techniques achieving higher resolution or reduced noise have thus been shown to improve segmentation accuracy [33]. These gains stem from superior visualization of internal architecture (e.g., strata radiatum) and pathology (e.g., sclerosis), confirming that image fidelity improvements intrinsically support clinical decision making. Dedicated studies with larger cohorts of confirmed hippocampal pathology are warranted to validate these relationships, including segmentation quality and accuracy for SRR, specifically.

Beyond optimizing sequence parameters and incorporating AI to enhance image quality, moving to higher field strength has been suggested as a powerful means of bringing hippocampal imaging to a new level. Specifically, the advent of 7T MRI for clinical purposes offers unprecedented potential. With a hippocampal sublayer thickness ranging from 0.1–0.5 mm, the resolution required to visualize these structures approaches the practical limits of in vivo full FOV MR imaging at lower field strengths, such as 1.5 or 3T [34]. The increased signal-to-noise ratio and spatial resolution afforded by 7T MRI allow for more detailed visualization of these delicate sublayers, providing new opportunities for advanced neuroimaging studies and improved understanding of hippocampal function and pathology [4,35]. Whether 3T MRI enhanced by AI can keep up with 7T remains to be seen and should be investigated in future studies.

Our study exhibits the following limitations: Firstly, this was a single-center study performed on a limited number of subjects. Secondly, our study employed hardware and software exclusively from a single vendor, limiting the generalizability of results to other platforms. While this inherent constraint introduces potential vendor-specific bias, we mitigated analytical influences through blinded assessment protocols and independent processing pipelines. Furthermore, we limited our study and analyses to a specific set of MRI acquisition and reconstruction parameters. We acknowledge that our results also inherently depend on the selection of these parameters. Thirdly, we limited our analysis to subjective image quality and subjective evaluation of anatomical structures. Diagnostic accuracy was not assessed due to the very limited number of subjects presenting with relevant/pathological imaging findings. Moreover, we did not evaluate objective image quality of the sequences. However, it should be noted that conventional ROI-based signal-to-noise ratio (SNR) metrics can be compromised by the inherent denoising algorithm of CS- and CS-AI-based technologies, where the regional distribution of noise within the image may inherently vary [36].

In conclusion, our findings demonstrate that DL-powered SSR can enhance imaging performance by introducing flexibility in T2w hippocampal imaging, either by improving image quality in non-accelerated scans or by maintaining acceptable quality in accelerated scans, with the added benefit of reduced scan times. Futures studies should evaluate the impact of SSR on diagnostic performance.