Introduction

In hepatic MRI, T2-weighted imaging facilitates the detection and characterization of hepatic lesions [1,2,3,4]. However, its long acquisition time causes numerous artifacts from the patient’s respiratory movements. The development of turbo spin echo (TSE) sequences with respiratory triggering or the use of free-breathing in combination with parallel imaging initially enabled reduced acquisition times and artifacts [5,6,7,8,9,10]. MR manufacturers have developed sequences using k-space sampling strategies to further reduce motion artifacts (PROPELLER at General Electric, BLADE at Siemens, MultiVane at Philips) [11,12,13,14,15,16]. Although the image quality has been significantly improved, the acquisition time is still too long in view of the increasing indications for MRI examinations [17, 18].

The half-Fourier acquisition single-shot turbo spin echo (HASTE) sequence is a single-shot sequence that is used to acquire all the raw imaging data needed for a slice in a few hundred milliseconds using a single echo train following a 90° excitation pulse. Thus, it is robust to motion, and related artifacts are therefore significantly reduced [19, 20]. Compared to a conventional T2-weighted sequence, a HASTE sequence uses a partial Fourier acquisition to ensure the desired echo time (TE) for center k-space lines, and uses longer echo trains.

Depending on the extent of partial Fourier sampling and the duration of the echo train, the image contrast can be affected. The T2 decay during the acquisition of an echo train leads to high T2 weighting at higher frequencies, and the resulting images suffer from a significant degradation of the signal-to-noise ratio (SNR) due to the weak intensity of the late echo signals [21]. The spatial resolution is also reduced with the presence of blurring in the direction of phase encoding [19]. To improve the image quality of the HASTE sequence and make the image contrast more comparable to that of conventional TSE imaging, a shorter echo train, less aggressive partial Fourier imaging, and a higher SNR have been used.

One avenue of research could be the use of a deep learning–based reconstruction associated with the HASTE sequence (DL HASTE), such as the sequence recently developed by one vendor, which would allow for higher acceleration and image acquisition with a better SNR. Moreover, it could allow a reduction in the echo train length due to an additional separate acquisition of the calibration data.

In previous retrospective studies, the DL HASTE sequence on abdominal MRI at 1.5 T has been compared with the T2-weighted BLADE sequence [22, 23] and the T2 TSE sequence [24]. Herrmann J et al studied the impact of flip angle variation on the image quality and diagnostic confidence and then retrospectively compared the DL HASTE sequence and T2-weighted BLADE sequence for a small number of patients for liver MRI at 3 T [25]. More recently, Ginocchio et al compared the DL HASTE sequence for liver MRI at 3 T with the conventional T2 TSE sequence [26].

To our knowledge, our study is the first to prospectively compare, qualitatively and quantitatively, the DL HASTE and T2-weighted BLADE sequences for liver MRI at 3 T. In addition, we quantitatively compare the two sequences by measuring the signal-to-noise ratio and the contrast-to-noise ratio, which has not yet been performed in other studies at 3 T.

The objective of this study is to qualitatively and quantitatively compare the DL HASTE sequence with the conventional T2-weighted BLADE for 3-T liver MRI acquisitions. We hypothesize that the DL HASTE sequence is neither qualitatively nor quantitatively inferior to the conventional T2-weighted BLADE sequence and has a much shorter acquisition time.

Methods

Study population

The AIMANT (optimization of the acquisition sequence to improve the quality and comfort of magnetic resonance imaging exam, NCT04645628) study was conducted in compliance with the Declaration of Helsinki and was approved by the institution’s Ethics Committee (number 2020-063B).

Consecutive adult patients referred in our institution for a clinically indicated liver MRI between December 2020 and January 2021 were included in this prospective, monocenter study. Missing patient data and incomplete MRI datasets were considered exclusion criteria.

All participants provided written informed consent.

MRI acquisitions

All MR examinations were performed using the same 3-T clinical MR scanner (MAGNETOM Skyra; Siemens Healthcare), using a body array coil with 18 channels and a spine array coil with 32 channels, with the patient in the dorsal decubitus position. For each patient, the MR imaging protocol included a fat-suppressed T2-weighted BLADE sequence with navigator-triggered prospective acquisition correction (PACE) in the axial plane and a fat-suppressed DL HASTE sequence with a single breath-hold in the axial orientation.

The acquisition parameters of both sequences are detailed in Table 1.

Table 1 Acquisition parameters of the T2-weighted BLADE and DL HASTE sequences on 3-T liver MRI
Full size table

The modifications applied to the prototypical DL HASTE sequence compared to the conventional HASTE sequence are aimed at shortening the echo train duration, reducing the specific absorption rate (SAR), reducing T2-blurring, and shortening the overall acquisition time. To shorten the duration of the echo train, a separate acquisition of the reference data for the estimation of coil sensitivity maps was performed. This was realized by a second echo train following the acquisition of the imaging data. The SAR was thereby reduced by having fewer refocusing pulses in the echo train and the use of a higher acceleration factor compared with that of the conventional sequence. Furthermore, a variable flip angle evolution for the refocusing pulses was used to reduce the SAR and delay the T2 relaxation. The evolution was determined by the initial flip angle (130°), the minimal flip angle (100°), the flip angle at the k-space center (120°), and the final flip angle (140°) [27]. The effective echo time (TE) was 99 ms, the apparent TE due to the delayed T2 decay was 79 ms, and the repetition time (TR) was set to 595 ms. Compared to the conventional HASTE sequence, the slice increment between sequentially acquired slices was set to 4 to mitigate signal attenuation due to crosstalk and magnetization transfer at such a small TR.

Other sequences analyzed in the complementary analysis included diffusion-weighted imaging (b values of 50, 500, and 800 s/mm2) performed in free-breathing using a simultaneous multislice echoplanar imaging (SMS-EPI) system. A T1-weighted VIBE sequence with fat suppression using spectral attenuated recovery in one breath-hold was also analyzed. After a precontrast acquisition, a standard dose (0.2 mL/kg) of gadoxetic acid (Clariscan; GE Healthcare) was intravenously injected at a rate of 2 mL/s by using a power injector (OptiStar Elite; Guerbet) followed by a 20 mL saline flush. Arterial phase images were acquired 20 s after contrast administration, and portal venous phase and delayed phase images were obtained 80 s and 3 min after contrast administration, respectively. Detailed information about these sequences is summarized in Supplemental Table 1.

Image reconstruction with deep learning HASTE

Image reconstruction of the DL HASTE sequence is achieve by using an unrolled iterative reconstruction network, where a gradient step that improves the data consistency is followed by a regularization step. This architecture shares some similarities with the variational network as introduced previously [28]. The network uses coil sensitivity maps that have been precomputed from a separate reference scan, k-space data, and bias-field correction.

This network uses two types of iterations; each of them has data consistency and trainable (Nesterov-type) extrapolation steps [29]. The first 22 iterations contain a trainable extrapolation and a data consistency step but no image regularization. The goal is to first minimize the data consistency and only focus on parallel imaging reconstruction. In the next 12 iterations, image regularization via a residual dense U-Net [30] is added after each data consistency step [24] to enhance the image quality.

Training data were acquired on volunteers using the conventional HASTE sequence. The imaging protocols were based on the vendor-supplied HASTE protocols for the abdomen in the axial and coronal orientations. The acquisitions were performed with an acceleration factor of 2, and the ground truth images required for the supervised training procedure were reconstructed using an unregularized SENSE reconstruction with the same estimated coil sensitivity maps as those used in the network reconstruction. As input to the network in the training phase, the k-space was then retrospectively downsampled to an acceleration factor of 4. In total, more than 10,000 slices were acquired for training. Compared to the protocols provided with the scanner, acquisition parameters were varied in the following ranges: TE from 81 to 200 ms, TR from 550 to 3800 ms, base resolution from 256 to 384, phase resolution from 70 to 95%, and phase oversampling from 0 to 50%. Acquisitions varied between unprepared, spectrally fat-suppressed, and inversion recovery prepared image contrasts. The acquisitions were performed on various clinical 1.5-T and 3-T scanners (MAGNETOM scanners, Siemens Healthcare) with a relative ratio of approximately 1:2.

Analysis of MRI images

MRI images were analyzed on a Picture Archiving and Communication System (PACS) viewing station (DirectView, version 12.2.2.1025, Carestream Health Inc.) by P.W., a radiology resident, and V.L., a radiologist with 15 years of experience in abdominal imaging. Each reviewer was blinded to the patient identification, clinical information, and possible biopsy results. Each sequence was assessed separately and without knowledge of the other sequences. Qualitative analysis and lesion identification of the T2-weighted BLADE and DL HASTE sequences were performed in two separate sessions, 1 month apart, to avoid possible recall bias. This review was performed for all participants for P.W. and for thirty randomly selected participants for V.L.

P.W. performed a third session 1 month later on all participants for the measurement of signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) in a joint unblinded analysis of the two sequences to allow for reproducibility of measurements and to avoid potential measurement bias.

P.W. performed a fourth session on all participants in a joint non-blinded complementary analysis with all liver MRI sequences to define true-positive and false-positive lesions.

Qualitative analysis

Qualitative analysis was performed subjectively with assessment of the overall quality of the sequence rated on a scale of 1 to 5 (1 = “mediocre,” 2 = “poor,” 3 = “moderate,” 4 = “good,” 5 = “excellent”) and presence of artifacts on a scale of 0 to 2 (0 = “no artifact,” 1 = “present but not interfering with interpretation,” 2 = “present and interfering with interpretation”). The reviewer then had to identify the smallest hepatic lesion larger than 5 mm and specify its conspicuity on a scale of 1 to 5 (1 = “mediocre,” 2 = “poor,” 3 = “moderate,” 4 = “good,” 5 = “excellent”) as well as its presumed nature on a scale of 0 to 3 (0 = “difficult to ascertain,” 1 = “liquid,” 2 = “solid,” 3 = “mixed”) and its location according to the Couinaud classification of liver segments [31].

Quantitative analysis

Identification of lesions

For each sequence, the total number of focal hepatic lesions (> 5 mm) and the size of the smallest visible focal hepatic lesion were recorded. Histopathologic confirmation was not obtained for lesions identified by investigators as focal hepatic lesions.

Signal-to-noise ratio and contrast-to-noise ratio measurements

To characterize the hepatic signal, a circular 20-mm-diameter region of interest (ROI) was placed in a healthy part of the right hepatic lobe while excluding hepatic vessels, lesions, any artifacts (repetition, metallic, etc.), and sudden signal changes as much as possible.

For the contrast measurement, maximum-sized circular ROIs were placed inside the two largest hepatic lesions identified by the investigator. The ROIs had to be placed within lesion edges and had to include all the components of the lesion if it was heterogeneous. For each ROI placed on a lesion, an identically sized ROI was placed in adjacent healthy parenchyma while excluding hepatic vessels, other lesions, any artifacts, and sudden signal changes as much as possible.

Each ROI was reproduced identically in terms of size and placement on the other investigated MRI sequence of the same patient in a joint unblinded analysis of the two sequences. The ROIs were replicated on the other sequence on synchronized slices, automatically or with manual adjustment, between the two sequences to allow the reproducibility of measurements.

In parallel imaging, noise is not evenly distributed across the image [21, 32]. Consequently, it cannot be measured globally and be accurately estimated by a measurement in a homogeneous region outside of the object of interest. Although an accurate measurement of noise across the entire image is impossible, a local estimate of the noise can be obtained. Therefore, in our study, we measured noise locally in a region close to the target signal measurement site, as suggested by Heverhagen [33].

The SNR was calculated according to the following formula:

$$\mathrm{SNR}= \frac{{\mathrm{SI}}_{\mathrm{liver}} }{{\mathrm{N}}_{\mathrm{liver}}}$$

where SIliver is the mean of the liver signal and Nliver is the standard deviation of the liver signal.

The CNR was calculated according to the following formula:

$$\mathrm{CNR}=\frac{({\mathrm{SI}}_{\mathrm{lesion}}- {\mathrm{SI}}_{\mathrm{adja}})}{{\mathrm{N}}_{\mathrm{adja}}}$$

where SIlesion is the mean of the signal from the hepatic lesion, SIadja is the mean of the signal from the adjacent healthy liver, and Nadja is the standard deviation of the signal from the adjacent healthy liver.

Complementary analysis

P.W. performed a complementary analysis on images of participants who had supernumerary liver lesions visible only on the DL HASTE sequence and not visible on the T2-weighted BLADE sequence or visible on the T2-weighted BLADE sequence and not visible on the DL HASTE sequence.

For each of these supernumerary lesions, their size and location according to the Couinaud classification were recorded. A joint nonblind analysis with the five other hepatic MRI sequences (diffusion-weighted imaging and T1-weighted VIBE: precontrast, arterial phase, venous phase, and delayed phase) allowed us to define whether the lesion was visible on these sequences.

The lesion is defined as visible when it appears in a hypersignal at any b value on diffusion-weighted imaging or when it appears in a hypersignal or hyposignal on T1 VIBE precontrast and/or T1 VIBE postcontrast sequences.

Statistical analysis

Continuous variables are expressed as the median ± interquartile range (IQR), and categorical variables are expressed as frequencies (percentages). The normality of the distribution was tested by using the Shapiro-Wilk test.

A paired Wilcoxon signed-rank test with continuity correction was used to compare the number of visible lesions, their sizes, and the CNR and SNR between the two sequences. The chi-square test was used to compare the quality, presence of artifacts, and conspicuity of the smallest lesion between the two sequences. McNemar’s test was used to compare the presumed nature of the smallest lesion between the two sequences (dichotomized into two groups: presumed nature judged as “difficult to ascertain” versus “liquid” or “solid” or “mixed”).

A multivariate logistic analysis was performed to determine a potential relationship between the quality of the examination (dichotomized into two groups: quality assessed as “good” or “excellent” versus “moderate” or “poor” or “mediocre”) and the following three explanatory variables: age, BMI, and presence of ascites.

A p value < 0.05 was considered to indicate a statistically significant difference.

Intraclass correlation coefficients (two-way random effects, absolute agreement) (ICCs) were used to assess the interreader agreement for quantitative variables such as the total number of focal liver lesions detected and the size of the smallest visible focal liver lesion. The weighted kappa (кw) values were used for the agreement of ordinal qualitative data such as the overall quality, presence of artifact, and conspicuity of the smallest lesion, and Cohen’s kappa (к) was used for the nominal qualitative data such as the presumed nature of the smallest lesion and its location.

ICCs or к coefficients between 0.81 and 1.00 indicate an excellent agreement; those between 0.61 and 0.80 indicate a strong agreement; those between 0.41 and 0.60 indicate a moderate agreement; those between 0.21 and 0.40 indicate a moderate agreement; and those between 0.00 and 0.20 indicate a poor agreement.

For complementary analyses, the size, location, and number of true-positives and false-positives of supernumerary lesions were calculated. A true-positive is defined as a lesion visible only on the DL HASTE or T2-weighted BLADE sequence and visible on at least one of the other five liver MRI sequences. A false-positive is defined as a lesion visible only on the DL HASTE or T2-weighted BLADE sequence and not visible on the five other sequences.

Statistical analysis was performed by G.H. using R software, version 4.0.3, R Project for Statistical Computing, including its CBCGrps library [34].

Results

Patient characteristics

We enrolled 114 patients in the AIMANT study. Of these, two patients were excluded because of missing MRI data. The data from a total of 112 participants were analyzed (Fig. 1). The mean age ± standard deviation was 62.5 years ± 13.9; 75 participants (67%) were men; the mean body mass index ± standard deviation was 27.9 kg/m2 ± 5.9. Table 2 summarizes the baseline characteristics of the participants.

Fig. 1
figure 1

Flowchart of patients through the AIMANT study. Abbreviations: HCC, hepatocellular carcinoma

Full size image
Table 2 Baseline characteristics of the study population
Full size table

Qualitative analysis

The results of the qualitative analysis are summarized in Table 3.

Table 3 Results of the qualitative analysis of the DL HASTE and T2-weighted BLADE sequences
Full size table

The quality of the examination was significantly higher with the DL HASTE sequence; 69% (77 cases) of the exams were assessed as having good to excellent image quality (≥ 4) using the DL HASTE sequence versus 44% (49 cases) with the T2-weighted BLADE sequence (p = 0.006; Fig. 2).

Fig. 2
figure 2

Characterization of focal liver lesions in a 68-year-old male. The DL HASTE axial image (a) shows better overall image quality and fewer motion artifacts than the T2-weighted BLADE axial image (b)

Full size image

The presence of artifacts was significantly lower for the DL HASTE sequence than for the T2-weighted BLADE sequence (p < 0.001), with only 9% (10 cases) of the examinations assessed with artifacts interfering with interpretation for the DL HASTE sequence against 18% (20 cases) for the T2-weighted BLADE sequence (Fig. 2).

The conspicuity of the smallest focal hepatic lesion was significantly better for the DL HASTE sequence, with 71% of the lesions assessed with good to excellent conspicuity (≥ 4) versus 39% with the T2-weighted BLADE sequence (p = 0.001; Fig. 3).

Fig. 3
figure 3

Characterization of focal liver lesions in a 59-year-old male followed for a renal cell carcinoma. The DL HASTE axial image (a) shows better conspicuity for a small 5-mm focal liver lesion (arrow) than the T2-weighted BLADE axial image (b). Other focal hepatic lesions, particularly those of the left liver, are also better visualized on the DL HASTE image

Full size image

The presence of ascites was identified as a potential explanatory variable, although no significant difference was observed between the two sequences in this subset of patients (p = 0.096).

Quantitative analysis

The results of the quantitative analysis are summarized in Table 4.

Table 4 Results of the quantitative analysis of the DL HASTE and T2-weighted BLADE sequences
Full size table

Significantly more lesions were identified using the DL HASTE sequence (median, 1; IQR, 0–4) than the T2-weighted BLADE sequence (median, 1; IQR, 0–3.25; p < 0.001). We identified 356 lesions on the DL HASTE sequence versus 320 lesions on the T2-weighted BLADE sequence (Fig. 4).

Fig. 4
figure 4

Postoperative control of a left liver lobectomy for angiomyolipoma in a 70-year-old female. The DL HASTE axial image (a) shows two focal hypersignals mimicking liver lesions (arrows) corresponding to an uncanceled intravascular signal from the hepatic veins. Those hypersignals are not visible on the axial T2-weighted BLADE image (arrowheads) (b)

Full size image

The SNR values of the liver parenchyma were significantly better for the T2-weighted BLADE sequence (median, 10.28; IQR, 8.1–12.3) than for the DL HASTE sequence (median, 9.41; IQR, 7.44–11.05; p < 0.001). The CNR values of liver lesions were, however, significantly better for the HASTE DL sequence (median, 17.11; IQR, 7.86–31.67) than for the T2-weighted BLADE sequence (median, 12.10; IQR, 6.13–20; p < 0.001; Fig. 5).

Fig. 5
figure 5

Characterization of focal liver lesions in a 38-year-old female followed for adrenocortical cancer. The DL HASTE axial image (a) shows better conspicuity of a segment IV liver lesion (arrow) than the T2-weighted BLADE axial image (b). Two identical regions of interest (ROIs) are placed within the lesion (ROIlesion) and in adjacent healthy liver (ROIadjacent) on the HASTE DL image (c). The values indicate the mean and standard deviation of the signal. d The procedure is repeated identically on the T2-weighted BLADE image. We found a higher lesion contrast-to-noise ratio (CNR) on the HASTE DL image (CNR, 15.9) than on the T2-weighted BLADE image (CNR, 8.8), based on the previously described calculation method

Full size image

Interreader agreement

The interreader agreement between the two radiologists for the DL HASTE sequence was excellent for the total number of focal liver lesions detected (ICC, 0.986) and the size of the smallest visible liver lesion (ICC, 0.997). It was strong to excellent for the overall quality (кw, 0.820), presence of artifacts (кw, 0.789), conspicuity of the smallest lesion (кw, 0.836), presumed nature (к, 0.65), and location (к, 0.94).

For the T2-weighted BLADE sequence, the interreader agreement was excellent for the total number of focal liver lesions detected (ICC, 0.974) and the size of the smallest visible liver lesion (ICC, 0.998). It was moderate to excellent for the overall quality (кw, 0.788), presence of artifacts (кw, 0.712), conspicuity of the smallest lesion (кw, 0.733), presumed nature (к, 0.58), and location (к, 0.94).

Complementary analysis

The results of the complementary analysis are summarized in Table 5.

Table 5 Results of the complementary analysis of the DL HASTE and T2-weighted BLADE sequences
Full size table

Of the 41 supernumerary lesions visible only on the DL HASTE sequence, 38 (93%) were true-positives that were also visible on the other MRI sequences, and 3 (7%) were false-positives. The true-positives were located mainly on segments 2 (11 lesions; 27%) and 8 (9 lesions; 22%).

All 5 supernumerary lesions visible only on the T2 BLADE-weighted sequence were true-positives.

Discussion

In this study, we assessed the qualitative and quantitative performances of the DL HASTE sequence compared to the conventional T2-weighted BLADE sequence for liver 3-T MRI. It was found that the image quality significantly improved, the presence of artifacts was reduced, the conspicuity of small lesions was improved, and a larger number of hepatic lesions could be detected using the DL HASTE sequence compared to using the T2-weighted BLADE sequence.

One possible explanation for these results is the fast speed of acquisition. Indeed, all slices of the sequence were obtained in 21 s in a single breath-hold.

This allowed us to obtain a robust and reproducible sequence with fewer artifacts—particularly motion artifacts [19, 20, 35]. Moreover, given the high number of focal hepatic lesions in a T2 hypersignal, the strong T2 weighting of the HASTE sequence linked to the filling of the majority of k-space with long-TE echoes [19, 20, 35] enabled excellent contrast to be obtained, thus resulting in greater conspicuity and detection of small lesions.

A better CNR was obtained using the DL HASTE sequence than the T2-weighted BLADE sequence, whereas a better SNR was obtained using the T2-weighted BLADE sequence. An increase in image noise and image blurring in single-shot sequences [20, 35] due to an important difference in signal intensity between two lines in the k-space during a long echo train may represent a possible explanation for the decrease in SNR. Another possible explanation could stem from training of the neural network on a relatively small number of patient data, inducing an insufficient increase in signal intensity on the late echoes. In addition, fat saturation of the DL HASTE sequence induces a decrease in overall signal intensity [19, 20]. Although signal intensity decreases globally in the image, the contrast-to-noise ratio is better with the DL HASTE sequence, which is a crucial element for hepatic lesion detection.

Of the 38 true-positives of supernumerary lesions visible only on the DL HASTE sequence, the great majority seemed to be explained by the presence of artifacts, notably motion artifacts on the T2-weighted BLADE sequence, which did not allow detection of these lesions. Moreover, this hypothesis seems to be supported by the fact that these lesions were preferentially localized in liver segments 2 and 8, which are subject to more artifacts, particularly those related to cardiac movements. Of the 3 false-positives with the DL HASTE sequence, two seemed to be explained by a signal cancelation defect within hepatic vessels. Indeed, both reviewers noted that in using the DL HASTE sequence, the intravascular signal was not canceled in some slices, particularly in the hepatic veins. These focal intravascular hypersignals could therefore potentially induce the detection of false-positives, especially when they are localized in only one slice. This artifact is identifiable, especially on large- and medium-caliber vessels, because it follows a vascular pathway and the hypersignal takes on a long form or dichotomous appearance (Fig. 4). In addition, the analysis of the overlying and underlying slices allows tracking of the vessel and confirming that the artifact is intravascular.

No explanation was found regarding the last of the three false-positives for the DL HASTE sequence. The five true-positives visible only on the T2-weighted BLADE sequence seemed to be explained by the absence of slice passing through these small lesions on the DL HASTE sequence. This could be explained by a bad breath-hold of the patient.

In a study in abdominal MRI at 1.5 T [22], Herrmann et al noted, in contrast to our study, a better image quality and diagnostic confidence of the T2-weighted BLADE sequence compared to the DL HASTE sequence. We used a higher magnetic field (3 T) to increase the signal-to-noise ratio and possibly the spatial resolution. The neural network of the DL HASTE sequence used in the 1.5-T MRI is different from the one at 3 T to increase denoising while also increasing blur. In addition, the spatial resolution was increased as a result of the larger matrix size we used.

Similarly, in the study by Mulé et al [23], the DL HASTE and T2-weighted BLADE sequences were compared for liver MRI at 1.5 T. Similar to our study, their results showed a better image quality and a better CNR using the DL HASTE sequence. Contrary to our results, the SNR was better using the DL HASTE sequence. A possible explanation could be the different calculation methods between our two studies.

Shanbhogue et al [24] also noted an improved overall image quality, a reduction in the presence of artifacts, and an improved lesion visibility using the DL HASTE sequence. This study compares the DL HASTE sequence with the conventional turbo spin echo sequence without the technique of oversampling the data bands in a rotational manner around the center of the k-space (PROPELLER technique), which improves the image quality and reduces the presence of artifacts. As previously described, this study used a lower magnetic field (1.5 T) and a smaller matrix size.

Another study compared the DL HASTE sequence with two different flip angle evolution patterns to the T2-weighted BLADE sequence for abdominal MRI at 3 T [25]. Contrary to our study, these authors noted a better quality, sharpness, and contrast and less noise when using the T2-weighted BLADE sequence than using the DL HASTE sequence. A possible explanation may be the larger matrix and longer TE of the DL HASTE sequence used in our study, which could improve the spatial resolution and T2 contrast. Conversely, the matrix size of our T2-weighted BLADE sequence was smaller than in that in their study.

More recently, Ginocchio et al found similar results to our study for liver MRI at 3 T for overall quality, artifact reduction, lesion conspicuity, and lesion detection for the DL HASTE sequence [26]. This study included a small sample of participants in comparison with the conventional T2-weighted TSE sequence with potentially worse image quality and more artifacts than a sequence using a PROPELLER technique. The matrix size of both sequences was also smaller than in our study.

Several limitations in our study must be mentioned.

First, given its single-center design, our study may suffer from potential selection bias.

Second, since both sequences compared in our study are based on parallel imaging, the estimation of both SNR and CNR is difficult, and the technique implemented in our study corresponds to a local estimation close to the region of interest. This approach has the advantage of being fast and easy to perform. Only a noise measurement from a phantom or acquiring the same image twice would enable measuring the overall noise of the image [21, 36].

Concerning the greater number of focal hepatic lesions detected using the DL HASTE sequence, we recall that no histopathological confirmation has been performed on these supernumerary lesions detected only on the DL HASTE sequence despite a joint analysis of the other MRI sequences.

Contrary to a previous abdominal MRI study at 1.5 T, we did not compare the DL HASTE sequence with the conventional HASTE sequence because this sequence is not performed in liver MRI at our center as a part of clinical practice. This would have been an interesting avenue of research to study the neural network modifications on the conventional HASTE sequence, especially in pancreatic MRI or cholangio-MRI, where this sequence is used in routine practice.

In conclusion, our findings show that the DL HASTE sequence is more robust to artifacts, particularly motion artifacts, than the T2-weighted BLADE sequence. It also features better image quality and excellent contrast, allowing a greater number of hepatic lesions to be detected than with the T2-weighted BLADE sequence.