Abstract
Objective
The objective of this study was to evaluate the clinical feasibility of deep learning reconstruction-accelerated thin-slice single-breath-hold half-Fourier single-shot turbo spin echo imaging (HASTEDL) for detecting pancreatic lesions, in comparison with two conventional T2-weighted imaging sequences: compressed-sensing HASTE (HASTECS) and BLADE.
Methods
From March 2022 to January 2023, a total of 63 patients with suspected pancreatic-related disease underwent the HASTEDL, HASTECS, and BLADE sequences were enrolled in this retrospectively study. The acquisition time, the pancreatic lesion conspicuity (LCP), respiratory motion artifact (RMA), main pancreatic duct conspicuity (MPDC), overall image quality (OIQ), signal-to-noise ratio (SNR), and contrast-noise-ratio (CNR) of the pancreatic lesions were compared among the three sequences by two readers.
Results
The acquisition time of both HASTEDL and HASTECS was 16 s, which was significantly shorter than that of 102 s for BLADE. In terms of qualitative parameters, Reader 1 and Reader 2 assigned significantly higher scores to the LCP, RMA, MPDC, and OIQ for HASTEDL compared to HASTECS and BLADE sequences; As for the quantitative parameters, the SNR values of the pancreatic head, body, tail, and lesions, the CNR of the pancreatic lesion measured by the two readers were also significantly higher for HASTEDL than for HASTECS and BLADE sequences.
Conclusions
Compared to conventional T2WI sequences (HASTECS and BLADE), deep-learning reconstructed HASTE enables thin slice and single-breath-hold acquisition with clinical acceptable image quality for detection of pancreatic lesions.
Keywords: Deep learning, Half-Fourier acquisition single-shot turbo spin echo, Pancreatic lesions, Compressed sensing
1. Introduction
Magnetic resonance imaging (MRI), with its excellent spatial resolution and absence of radiation, plays a vital role in the diagnosis of pancreatic diseases[1]. T2-weighted imaging (T2WI), including fat-suppressed T2WI, as an essential component of pancreatic MRI sequences, enabling the detection and diagnosis of pancreatic structures and lesions [2]. However, conventional T2-weighted turbo spin echo (TSE) sequences are time-consuming (several minutes) and prone to motion artifacts due to multiple breath-hold and respiratory-triggered acquisitions[3]. These motion or respiratory artifacts can undermine radiologists’ confidence and accuracy in pancreatic diagnosis[4]. Furthermore, the pancreas, being thinner and surrounded by adjacent tissues, requires higher image quality compared to other abdominal organs[5,6]. Image fuzziness, misregistration, reduced signal-to-noise ratio, or contrast can obscure significant pancreatic lesions and anatomical structures.
Despite the development of additional techniques such as respiratory-triggering or k-space filling techniques, which have somewhat reduced respiratory motion artifacts, they still require long image acquisition time and multiple breath-holds[7,8]. Single-shot T2WI techniques, such as single-shot fast spin echo (single-shot FSE) and the half-Fourier acquisition single-shot turbo spin echo (HASTE), have been considered as alternatives for reduce motion artifacts due to their faster acquisition times and improved motion stability at baseline[9,10]. However, compared to conventional T2W TSE, the imaging contrast and quality in single-shot FSE T2WI techniques exhibit reduced imaging contrast and quality due to longer echo training [11]. In contrast, HASTE T2WI acquired less k-space data, allowing for shorter scan times and further reduction of motion artifacts at the cost of reduced spatial resolution and lesion conspicuity.
More acceleration methods have been developed and applied in the clinic to obtain T2W images with shorter acquisition time and higher quality [12,13]. For example, compression sensing (CS), an acceleration technique, has been employed to achieve faster scanning with improved image quality by bypassing the Nyquist-Shannon sampling criterion and utilizing non-coherent k-space sampling[14], [15], [16]. However, despite advancements in sparse transform algorithms, the reconstruction of undamaged and clear images from randomly and highly under-sampled k-space data, and the preservation of microscopic anatomical structures, remains challenging when utilizing the CS acceleration technique. Additionally, the incorporation of trainable deep learning (DL) algorithms, such as convolutional neural networks (CNNs), into reconstruction techniques to improve the inherent defects of conventional imaging scans has been intensively developed [17,18]. For instance, in our center, several modified T2WI based on DL-based reconstruction, such as single-breath-hold T2WI and artificial intelligence-assisted compressed sensing single-shot fluid-attenuated inversion recovery, have successfully applied in the liver and brain lesion detection[19,20]. DL reconstruction-accelerated HASTE (HASTEDL) has shown its advantages in abdominal MRI[21,22]. However, there is a lack of studies on DL techniques specifically applied to pancreas T2WI.
Therefore, the objective of our study was to evaluate the clinical feasibility of thin-slice single-breath-hold HASTEDL for the detection of pancreatic lesions and compare its performance with HASTE based on compressed sensing (HASTECS) and BLADE techniques.
2. Materials and methods
This retrospective study was approved by the Institutional Review Board of our hospital (approval no. B2020–346R), and the informed consent was waived. The process in our study strictly adheres to ethical- and hospital-related guidelines and regulations.
2.1. Participants
From March 2022 to April 2023, 63 patients with the suspected pancreatic-related disease who underwent abdominal 3T MRI were enrolled.
The inclusion criteria of this study were as follows: a) patients with radiologically assessable pancreatic lesions, with or without surgical or endoscopic puncture pathology confirmation; b) all patients with complete clinical and radiological data, including images from at least three sequences of HASTEDL and HASTECS, and BLADE. Additionally, the exclusion criteria consisted of: a) patients with claustrophobia or unable to complete the examination, b) poor image quality due to severe metallic or motion artifacts.
2.2. Image acquisition
All patients in the study underwent MRI using a 3-T MR scanner (MAGNETOM Prisma; Siemens Healthcare, Erlangen, Germany). The examination protocol consisted of conventional T1WI, diffusion-weighted imaging (DWI, b = 0, 800 s/mm2), BLADE, and the research applications HASTEDL and HASTECS, which were randomly conducted before contrast. Then, dynamic enhanced imaging was conducted using gadolinium diethylenetriamine pentaacetic acid (Gd-DTPA, Bayer HealthCare, dose, 0.1 mmol/kg; rate, 2 mL/s). The arterial, portal, and delayed phases were obtained at 20, 90, and 180 s, respectively. In this study, the total scan time was 102 s for BLADE with four breath-holds and 16 s for both HASTEDL and HASTECS with single breath-hold. The rest of the detailed protocols of HASTEDL, HASTECS, and BLADE are listed in Table 1.
Table 1.
Different s sequences protocols.
| HASTEDL | HASTECS | BLADE | |
|---|---|---|---|
| TR (ms) | 628 | 628 | 3000 |
| TE (ms) | 105 | 105 | 114 |
| Matrix | 320×320 | 320×320 | 320×320 |
| Field of view (mm2) | 360×360 | 360×360 | 360×360 |
| Voxel size (mm3) | 1.1 × 1.1 × 4 | 1.1 × 1.1 × 4 | 1.1 × 1.1 × 4 |
| Section thickness (mm) | 4 | 4 | 4 |
| Slice gap (mm) | 0.4 | 0.4 | 0.4 |
| Number of sections | 26 | 26 | 26 |
| Flip angle (◦) | 130 | 130 | 120 |
| Bandwidth (Hz) | 710 | 710 | 600 |
| Acceleration | 3.3 | 3.3 | 2 |
| Respiratory control | single breath-hold | single breath-hold | four breath-holds |
| Single breath-hold time (s) | 16 | 16 | 18 |
| Fat suppression | SPAIR | SPAIR | SPAIR |
| TA (s) | 16 | 16 | 102 |
TR, Repetition time; TE, Echo time; PAT, parallel acquisition technique; SPAIR, spectral-attenuated inversion recovery; TA, acquisition time.
2.3. MRI image analyses
All T2WI images acquired with HASTEDL, HASTECS, and BLADE were independently analyzed by two radiologists (Reader 1, X.X.X, and Reader 2, X.X.X, with 8 and 23 years of radiological experiences in pancreatic diseases, respectively) who were blinded to clinical, pathological, and radiological results during the review processing.
2.3.1. Qualitative imaging analysis
The pancreatic lesion conspicuity (LCP), respiratory motion artifact (RMA), main pancreatic duct conspicuity (MPDC), and overall image quality (OIQ) were independently and subjective scored by three readers for three T2WI sequences. These indicators were evaluated according to a five-point Likert scale as follows: 1 (severe): the presence of blurry edges of pancreatic lesions or structures, heavy image artifacts, and poor image continuity that do not meet the diagnostic requirements; 2 (moderate): the presence of relatively blurry edges of pancreatic lesions or structures, heavy image artifacts, and poor image continuity that make it difficult to meet the diagnostic requirements; 3 (mild): the presence of slightly blurry edges of pancreatic lesions or structure, and acceptable image artifacts and continuity that meet the basic diagnostic requirements. 4 (fine): the presence of relatively sharp and clear edges of pancreatic lesions or structure, and negligible image artifacts and continuity that meet the diagnostic requirements; 5 (excellent): pancreatic lesions or structures with completely sharp and clear edges, no image artifacts, good continuity, and fully compliant with diagnostic requirements. Moreover, slice continuity, defined as inconsistent horizontal position between layers, was divided into discontinuity or no discontinuity. The above two radiologists analyzed the slice continuity independently. When inconsistencies arose, a third radiologist with 25 years of radiological experience was asked to evaluate and use the unified opinion for follow-up analysis.
2.3.2. Quantitative image analysis
The signal intensities (SI) of the head (H), body (B), tail (T), and lesion (L) of the pancreas and the SI of the erector spinal muscle (ESM) were measured on operator-defined regions-of-interest (ROI) on the three T2WI images, respectively (Fig. 1). The sizes of ROI in the vertical spine muscle and pancreatic tissue and lesions are about 50∼100 mm2, and the corresponding ROI is placed in the signal uniformity area, avoiding the influence of blood vessels and bones. The image signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) are calculated according to the following formulas:
SNR of pancreatic head: SNRH =SIH / SIESM;
SNR of pancreatic body: SNRB =SIB / SIESM;
SNR of pancreatic tail: SNRT =SIT / SIESM;
SNR of pancreatic lesions: SNRL =SIL / SIESM;
CNR of pancreatic lesions: CNRL=|SIL-SIESM| / SIESM
Fig. 1.
Representative ROI placement for the head (1), body (2), tail (3), and lesion (4) of the pancreas.
2.4. Statistical analysis
Statistical analyses were performed using SPSS version 26 (IBM Corp., Armonk, NY, USA). The image quality scores obtained for three T2WI sequences were compared using a paired Wilcoxon signed-rank test. P < 0.05 was considered as statistically significant and Bonferroni correction was conducted to correct for cumulative α error [23]. Interobserver agreement was assessed using the Cohen κ coefficient [24]. The κ values were used to interpret according to the following criteria: poor (0–0.20), fair (0.21–0.40), moderate (0.41–0.60), good (0.61–0.80), almost perfect (0.81–1).
3. Results
3.1. Patient characteristics
A total of 63 patients (mean age, 61.13 ± 11.25 years; range, 30–83 years, 34 men and 29 women;) with pancreatic lesions were enrolled in the present study for subsequent analysis.
Regarding pancreatic lesions, surgical tumor resection with pathological confirmation was performed on 31 patients. Among them, 23 patients were diagnosed with pancreatic ductal adenocarcinoma (PDAC), while 8 patients had pancreatic neuroendocrine tumors (G1, n = 2 and G2, n = 6). Additionally, 11 patients received pathologically confirmed through endoscopic puncture and were diagnosed PDAC. The remaining patients were diagnosed based on the indications of abdominal MRI and clinical signs and symptoms, including pancreatic solid tumor (n = 7), pancreatic cystic lesions (n = 12) and suspected pancreatic inflammatory disease (n = 2). The size of the lesions (maximum diameter) ranged from 1.1 to 7.1 cm, with a mean of (2.9 ± 1.4) cm. For the location of pancreatic lesions, there were 36 in the head of pancreas, 12 in the body of pancreas, and 19 in the tail of pancreas. Moreover, 11 cases (17.5 %) in HASTEDL showed slice discontinuities, significantly less than 18 (28.6 %) and 20 (31.7 %) cases in HASTECS and BLADE.
3.2. Qualitative imaging analysis
The three T2WI protocols were successfully examined and analyzed in all patients. The acquisition time (AT) for both HASTEDL and HASTECS was 16 s, while the AT for BLADE was 102 s.
The five-point Likert scale was used to assess four qualitative image qualities (LCP, RMA, MPDC and OIQ) for HASTEDL, HASTECS, and BLADE sequences (Tables 2 and 4). In terms of pancreatic lesion analysis, a total of 65 pancreatic lesions were detected and evaluated. The LCP, RMA, and OIQ scores of HASTEDL in Reader 1 and Reader 2 were all significantly higher than those for HASTECS and BLADE sequences (P value ranging from <0.001 to 0.003) (Fig. 2, Fig. 3). However, there were no statistical differences in these scores between HASTECS and BLADE. For pancreatic ducts, HASTEDL achieved highest MPDC scores of 4.67±0.48 and 4.68±0.47 in reader 1 and reader 2, respectively, which were significantly higher than those of HASTECS and BLADE (P value ranging <0.001 to 0.006) (Fig. 4). In terms of MPDC, the scores for HASTECS in both Reader 1 and Reader 2 were 4.48±0.59 and 4.49±0.56, respectively, which were higher than those of BLADE.
Table 2.
The qualitative parameters among HASTEDL, HASTECS and BLADE with kappa values.
| Reader 1 |
Reader 2 |
Kappaa | Kappab | Kappac | |||||
|---|---|---|---|---|---|---|---|---|---|
| HASTEDL | HASTECS | BLADE | HASTEDL | HASTECS | BLADE | HASTEDL | HASTECS | BLADE | |
| LCP | 4.60±0.43 | 4.02±0.66 | 4.05±0.88 | 4.75±0.44 | 4.00±0.65 | 4.03±0.89 | 0.872 | 0.917 | 0.930 |
| MPDC | 4.67±0.48 | 4.48±0.59 | 3.90±0.78 | 4.68±0.47 | 4.49±0.56 | 3.89±0.74 | 0.892 | 0.814 | 0.866 |
| RMA | 4.79±0.41 | 4.49±0.59 | 4.48±0.73 | 4.76±0.43 | 4.48±0.64 | 4.44±0.78 | 0.817 | 0.855 | 0.885 |
| OIQ | 4.63±0.49 | 4.02±0.46 | 3.89±0.81 | 4.62±0.49 | 3.97±0.51 | 3.83±0.75 | 0.831 | 0.763 | 0.844 |
LCP: pancreatic lesion conspicuity; RM: respiratory motion artifact; MPDC: main pancreatic duct conspicuity; OIQ: overall image quality; a Inter-reader agreement for HASTEDL image was evaluated by using Cohen's kappa; b Inter-reader agreement for HASTECS image was evaluated by using Cohen's kappa; c Inter-reader agreement for BLADE image was evaluated by using Cohen's kappa;.
Table 4.
The comparisons of qualitative and quantitative parameters among HASTEDL, HASTECS and BLADE.
| Imaging quality | Reader 1 | Reader 2 | ||||
|---|---|---|---|---|---|---|
| PDLvsCS | PDLvsBL | PCSvsBL | PDLvsCS | PDLvsBL | PCSvsBL | |
| LCP | <0.001 | <0.001 | 0.809 | <0.001 | <0.001 | 0.809 |
| MPDC | 0.006 | <0.001 | <0.001 | 0.004 | <0.001 | <0.001 |
| RMA | <0.001 | 0.002 | 0.834 | 0.001 | 0.003 | 0.582 |
| OIQ | <0.001 | <0.001 | 0.185 | <0.001 | <0.001 | 0.129 |
| SNRH | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
| SNRB | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
| SNRT | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
| SNRL | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
| CNRL | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 | <0.001 |
LCP: pancreatic lesion conspicuity; RMA: respiratory motion artifact; SNRH: SNR of pancreatic head; SNRB: SNR of pancreatic body; SNRT: SNR of pancreatic tail; SNRL: SNR of pancreatic lesions; CNRL: CNR of pancreatic lesions; MPDC: main pancreatic duct conspicuity; OIQ: overall image quality; P value less than 0.017 after Bonferroni correction are statistical significance.
Fig. 2.
A 66-year-old male with a pancreatic head tumor. The pancreas and pancreatic lesion (the arrow) can be clearly shown in HASTEDL (a). HASTECS (b) also showed the pancreas and pancreatic lesion (the arrow); however, the contrast between lesion and the pancreas was slightly poor. The lesion shape displayed in BLADE sequence (c) is fuzzy, and the contrast is poor; The contrast-enhanced T1WI sequence (d), the gold standard for diagnosis (the portal phase), shows the pancreatic head tumor (the arrow).
Fig. 3.
An 80-year-old male with pancreatic ductal adenocarcinoma in the tail of pancreas. The pancreas and pancreatic lesion (the arrow) can be clearly shown in HASTEDL (a) with low noise and high SNR, significantly better than those in HASTECS (b) and BLADE (c); The contrast-enhanced T1WI sequence (d), the gold standard for diagnosis (the portal phase), shows the pancreatic lesion (the arrow).
Fig. 4.
A 71-year-old female with cystic lesions of the tail of the pancreas. The pancreas and pancreatic lesion (the arrow) can be clearly shown in HASTEDL (a), HASTECS (b), and BLADE (c); Peripancreatic regional SNR performed best in HASTEDL, but the noise around the pancreas is higher in HASTECS and BLADE, decreasing the peripancreatic regional SNRs; Pancreatic duct (the triangular arrow) showed well in HASTEDL and HASTECS; however, BLADE offers poorly for the pancreatic duct; The contrast-enhanced T1WI sequence (d), the gold standard for diagnosis (the portal phase), shows the pancreatic lesion (the arrow).
The interobserver agreement of the qualitative parameters was good to almost perfect for HASTEDL (k = 0.817–0.892), HASTECS (k = 0.763–0.917), BLADE (k = 0.844–0.930).
3.3. Quantitative image analysis
In terms of the anatomical structures of the pancreas, the SNRH, SNRB, and SNRT values measured by both two readers were highest for HASTEDL, followed by HASTECS and BLADE; and these differences were statistically significant (Tables 3 and 4). Regarding to pancreatic lesions, CNRL and SNRL for HASTEDL in both readers were higher compared to HASTECS and BLADE.
Table 3.
The quantitative parameters among HASTEDL, HASTECS and BLADE with ICC values.
| Reader 1 | Reader 2 | ICCa | ICCb | ICCc | |||||
|---|---|---|---|---|---|---|---|---|---|
| HASTEDL | HASTECS | BLADE | HASTEDL | HASTECS | BLADE | HASTEDL | HASTECS | BLADE | |
| SNRH | 39.17±44.44 | 36.97±51.63 | 14.98±12.12 | 43.09±37.92 | 35.15±46.70 | 16.11±11.73 | 0.876 | 0.882 | 0.889 |
| SNRB | 37.66±43.06 | 35.91±51.27 | 14.78±12.66 | 41.94±46.97 | 32.59±46.16 | 16.09±14.03 | 0.891 | 0.850 | 0.757 |
| SNRT | 37.00±43.92 | 34.47±48.61 | 12.78±11.00 | 39.35±53.31 | 36.40±57.94 | 13.78±13.52 | 0.799 | 0.861 | 0.828 |
| SNRL | 90.62±103.74 | 83.74±128.75 | 34.52±25.64 | 102.17±108.50 | 89.34±131.22 | 39.01±39.74 | 0.893 | 0.825 | 0.755 |
| CNRL | 73.83±99.32 | 54.11±76.59 | 33.72±39.01 | 79.07±93.30 | 52.88±64.05 | 35.40±31.74 | 0.892 | 0.841 | 0.783 |
ICC, intraclass correlation; SNRH: SNR of pancreatic head; SNRB: SNR of pancreatic body; SNRT: SNR of pancreatic tail; SNRL: SNR of pancreatic lesions; CNRL: CNR of pancreatic lesions; a Inter-reader agreement for HASTEDL image was evaluated by using ICC; b Inter-reader agreement for HASTECS image was evaluated by using ICC; c Inter-reader agreement for BLADE image was evaluated by using ICC.
The interobserver agreement for the quantitative parameters was good to almost perfect for HASTEDL (ICC = 0.799–0.893), HASTECS (ICC = 0.825–0.882), BLADE (ICC = 0.755–889) (Table 3).
4. Discussion
In the present study, the feasibility of thin-slice single-breath-hold T2W HASTE with deep-learning-based reconstruction for detecting pancreatic lesions was investigated compared to BLADE and HASTECS. As results, the qualitative analyses demonstrated that the HASTEDL sequence exhibited improved pancreatic lesion conspicuity, main pancreatic duct conspicuity, overall image quality and less respiratory motion artifact compared to HASTECS and BLADE sequences. Regarding quantitative analyses, the HASTEDL sequence outperformed the other two sequences in term of SNR and CNR for both pancreatic structures and lesions.
The T2WI sequence plays a crucial role in the diagnosing and exploring pancreatic diseases [25]. Compared to other abdominal organs, the pancreas is relatively thin, located deeper in the body, and more susceptible to respiratory motion artifacts, therefore, fast thin-slice T2WI scanning is preferred for detecting pancreatic disease. Although thin-slice scan allows for higher spatial resolution, it can lead to a reduction in proton signal and SNR [26,27]. Therefore, effectively reducing acquisition time and minimizing the effects of motion artifacts and thin-slice scans are critical challenges in pancreatic T2WI imaging. In our study, both the HASTEDL and HASTECS sequences were able to complete a single breath-hold scan in 16 s. In contrast, the conventional BLADE sequence requires multiple breath-holds and had an acquisition time of 106 s, significantly increasing the time and cost required for routine clinical examinations.
The irregular morphology of the pancreas, its narrow alignment, and its central location within the abdomen make it more challenging to obtain clear images with clinically acceptable SNR compared to organs such as the liver closer to the coils [26]. Regarding the conspicuity of the pancreatic duct and lesions, HASTEDL was significantly preferred over the other two sequences, resulting in a substantial improvement in diagnostician confidence. Furthermore, the HASTEDL sequence demonstrated distinct advantages in suppressing respiratory artifacts and enhancing overall image quality. The conventional BLADE technique was initially developed to reduce motion artifacts in T2WI. However, the acquisition time of the BLADE sequence remains long. In our study, the acquisition time of BLADE was approximately six times higher than the other two sequences. While respiratory-triggered sampling methods theoretically allow for obtaining high-quality images, the effectiveness and image quality heavily rely on the patient's respiratory rhythm and amplitude. Unfortunately, most patients with pancreatic disease struggle to maintain a consistent and regular respiratory rhythm, resulting in variable acquisition times ranging from approximately 2 to 6 min. An additional concern is that when a patient's respiratory rhythm is unstable, techniques reliant on respiratory triggering may still produce respiratory artifacts and misaligned images, posing a particular risk for the inherently thin pancreas. This can increase the likelihood of losing images of the pancreatic lesion, thereby reducing diagnostic confidence.
DL has made significant advancements in recent years, integrating trainable deep learning algorithms into reconstruction techniques[28], [29], [30]. This integration addresses the limitations of conventional imaging scans and has been extensively developed for clinical applications [31]. DL reconstruction methods have demonstrated the ability to effectively reduce imaging noise while preserving image resolution[22,32] . In our present study, we employed regular under-sampling without integrated reference data in our thin-slice single-breath-hold HASTE T2WI, using a DL-based reconstruction approach. This approach aimed to shorten the echo train time and reduce the specific absorption rate, while reference data was obtained using a smaller flip angle, effectively reducing the specific absorption rate. Additionally, we incorporated variational networks that included bias-field correction and coil-sensitivity maps, along with iterations and trainable extrapolation steps. These techniques allowed for acquisitions without integrated calibration and offered flexibility in k-space sampling. Furthermore, to improve image quality, we utilized a residual dense U-net as image regularization [21]. The introduction of DL in HASTE retains its fundamental advantages of fast imaging and motion artifact reduction while significantly reducing noise and enhancing image resolution.
In our study, we also explored the application of CS for HASTE sequences in pancreas T2WI. CS-accelerated T2WI has gained popularity in the examination of abdominal organs, particularly the liver [33]. Previous studies on liver imaging have demonstrated that multiple breath-hold HASTECS sequences improve the signal-to-noise ratio and image quality compared to conventional multiple breath-hold T2WI [33]. Unfortunately, there is limited research comparing CS-accelerated T2WI and DL-accelerated T2WI imaging for pancreatic diseases [34]. We observed that the HASTECS sequences exhibited improved signal-to-noise ratio compared to BALDE. However, there was no significant difference in the objective evaluation of image quality. Regrettably, the CS-accelerated HASTECS sequence resulted in increased image noise, and the combination of high under-sampling and thin-slice scanning further exacerbated the reduced SNR of the images. The image SNR, image quality, and lesion visualization in HASTECS were inferior to those of HASTEDL. Theoretical analysis suggests that HASTECS struggles to reconstruct uncorrupted images from randomly and highly under-sampled k-space data or microanatomical structures, due to the limitations of sparse transform algorithms [35]. Additionally, compared to DL algorithms, the typical regularization and iterative optimization algorithms used in CS are simple and crude, lacking the trainability offered by DL. Considering the overall image quality and acquisition time, the single breath-hold HASTEDL demonstrated better application and robustness in fast thin-slice T2W imaging for the detection of pancreatic lesions.
However, there are several limitations in our study. Firstly, although we included a wide range of patients with various pancreatic diseases including tumors or inflammation, the sample size was relatively small which may introduce selection bias to some extent. Secondly, due to the timing and nature of the experiment, we were unable to obtain pathological results for a proportion of patients diagnosed with benign tumors or inflammatory lesions. Instead, their diagnoses relied on clinical, radiological and laboratory findings.
In conclusion, the thin-slice single-breath-hold HASTEDL demonstrates promising capability, providing significantly superior image quality and SNR for pancreatic lesions, compared with compared to HASTECS and BLADE. Moreover, HASTEDL exhibits excellent pancreatic lesion conspicuity and contrast, leading to increased diagnostic confidence among radiologists. Considering acquisition time and overall image qualities, HASTEDL could potentially become a standard protocol for thin-slice pancreatic MRI and may substitute conventional T2WI sequences.
Author contribution
All authors attest that they meet the current International Committee of Medical Journal Editors (ICMJE) criteria for authorship. The author(s) declare(s) that they had full access to all of the data in this study and the author(s) take(s) complete responsibility for the integrity of the data and the accuracy of the data analysis.
Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
Haitao Sun reports a relationship with Zhongshan Hospital Fudan University that includes: non-financial support.
Acknowledgments
Funding
This project was funded by "Science and Technology Innovation Action Plan" Star Cultivation (Sailing Program) (22YF1443600) and the China Postdoctoral Science Foundation (2023M732230).
Ethical approval
All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
References
- 1.Chu L.C., Goggins M.G., Fishman E.K. Diagnosis and detection of pancreatic cancer. Cancer J. 2017;23:333–342. doi: 10.1097/PPO.0000000000000290. [DOI] [PubMed] [Google Scholar]
- 2.Zhang B.B., Hou X.M., Chen Y.Q., Huo J.W., Jin E.H. Imaging features and risk factors of pancreatic cystic lesions complicating autoimmune pancreatitis: a retrospective study. Curr Med Imaging. 2023 doi: 10.2174/1573405619666230314095508. [DOI] [PubMed] [Google Scholar]
- 3.Glaser C., D'Anastasi M., Theisen D., Notohamiprodjo M., Horger W., Paul D., et al. Understanding 3d tse sequences: advantages, disadvantages, and application in MSK imaging. Semin Musculoskelet Radiol. 2015;19:321–327. doi: 10.1055/s-0035-1563732. [DOI] [PubMed] [Google Scholar]
- 4.Kabasawa H., Kiryu S. Pulse sequences and reconstruction in fast mr imaging of the liver. Magn Reson Med Sci. 2023;22:176–190. doi: 10.2463/mrms.rev.2022-0114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Chang H., Bates D., Gupta A., LeBedis C.A. Use of MR in pancreaticobiliary emergencies. Magn Reson Imaging Clin N Am. 2022;30:479–499. doi: 10.1016/j.mric.2022.04.009. [DOI] [PubMed] [Google Scholar]
- 6.Briola C. Magnetic resonance imaging and magnetic resonance imaging cholangiopancreatography of the pancreas in small animals. Vet Sci. 2022;9 doi: 10.3390/vetsci9080378. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Kraus M.S., Coblentz A.C., Deshpande V.S., Peeters J.M., Itriago-Leon P.M., Chavhan G.B. State-of-the-art magnetic resonance imaging sequences for pediatric body imaging. Pediatr Radiol. 2022 doi: 10.1007/s00247-022-05528-y. [DOI] [PubMed] [Google Scholar]
- 8.Chhetri A., Li X., Rispoli J.V. Current and emerging magnetic resonance-based techniques for breast cancer. Front Med. 2020;7:175. doi: 10.3389/fmed.2020.00175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Tyrell D., Mohamed F., Pavlides C., Kutalek S., Mulhern C., Nunes L.W. Half-fourier acquisition single-shot turbo spin echo imaging in the diagnosis of morgagni hernia. J Magn Reson Imaging. 2001;14:653–657. doi: 10.1002/jmri.1231. [DOI] [PubMed] [Google Scholar]
- 10.Castillo M., Mukherji S.K. Clinical applications of flair, haste, and magnetization transfer in neuroimaging. Semin Ultrasound Ct Mr. 2000;21:417–427. doi: 10.1016/s0887-2171(00)90034-9. [DOI] [PubMed] [Google Scholar]
- 11.Kabakus I.M., Spampinato M.V., Knipfing M., Cervantes G., Patel A., Eskandari R., et al. Fast brain magnetic resonance imaging with half-fourier acquisition with single-shot turbo spin echo sequence in detection of intracranial hemorrhage and skull fracture in general pediatric patients: preliminary results. Pediatr Emerg Care. 2021;37:e1168–e1172. doi: 10.1097/PEC.0000000000001949. [DOI] [PubMed] [Google Scholar]
- 12.Rajiah P.S., François C.J., Leiner T. Cardiac MRI: state of the art. Radiology. 2023;307 doi: 10.1148/radiol.223008. [DOI] [PubMed] [Google Scholar]
- 13.Zibetti M., Menon R.G., de Moura H.L., Zhang X., Kijowski R., Regatte R.R. Updates on compositional MRI mapping of the cartilage: emerging techniques and applications. J Magn Reson Imaging. 2023 doi: 10.1002/jmri.28689. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Taron J., Weiss J., Notohamiprodjo M., Kuestner T., Bamberg F., Weiland E., et al. Acceleration of magnetic resonance cholangiopancreatography using compressed sensing at 1.5 and 3 t. Invest Radiol. 2018;53:681–688. doi: 10.1097/RLI.0000000000000489. [DOI] [PubMed] [Google Scholar]
- 15.Zhu L., Xue H., Sun Z., Qian T., Weiland E., Kuehn B., et al. Modified breath-hold compressed-sensing 3D MR cholangiopancreatography with a small field-of-view and high resolution acquisition: clinical feasibility in biliary and pancreatic disorders. J Magn Reson Imaging. 2018;48:1389–1399. doi: 10.1002/jmri.26049. [DOI] [PubMed] [Google Scholar]
- 16.Hausmann D., Niemann T., Kreul D., Nocito A., Klarhöfer M., Nickel D.M., et al. Free-breathing dynamic contrast-enhanced imaging of the upper abdomen using a cartesian compressed-sensing sequence with hard-gated and motion-state-resolved reconstruction. Invest Radiol. 2019;54:728–736. doi: 10.1097/RLI.0000000000000607. [DOI] [PubMed] [Google Scholar]
- 17.Hu F., Chen A.A., Horng H., Bashyam V., Davatzikos C., Alexander-Bloch A., et al. Image harmonization: a review of statistical and deep learning methods for removing batch effects and evaluation metrics for effective harmonization. Neuroimage. 2023;274 doi: 10.1016/j.neuroimage.2023.120125. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Berbís M.A., Paulano G.F., Royuela D.V.J., Alcalá M.L., Luna A. Clinical impact of artificial intelligence-based solutions on imaging of the pancreas and liver. World J Gastroenterol. 2023;29:1427–1445. doi: 10.3748/wjg.v29.i9.1427. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Sheng R., Zheng L., Jin K., Sun W., Liao S., Zeng M., et al. Single-breath-hold T2WI liver MRI with deep learning-based reconstruction: a clinical feasibility study in comparison to conventional multi-breath-hold T2WI liver MRI. Magn Reson Imaging. 2021;81:75–81. doi: 10.1016/j.mri.2021.06.014. [DOI] [PubMed] [Google Scholar]
- 20.Liu K., Xi B., Sun H., Wang J., Chen C., Wen X., et al. The clinical feasibility of artificial intelligence-assisted compressed sensing single-shot fluid-attenuated inversion recovery (acs-ss-flair) for evaluation of uncooperative patients with brain diseases: comparison with the conventional t2-flair with parallel imaging. Acta Radiol. 2023;64:1943–1949. doi: 10.1177/02841851221139125. [DOI] [PubMed] [Google Scholar]
- 21.Herrmann J., Nickel D., Mugler J.P., Arberet S., Gassenmaier S., Afat S., et al. Development and evaluation of deep learning-accelerated single-breath-hold abdominal haste at 3 t using variable refocusing flip angles. Invest Radiol. 2021;56:645–652. doi: 10.1097/RLI.0000000000000785. [DOI] [PubMed] [Google Scholar]
- 22.Herrmann J., Gassenmaier S., Nickel D., Arberet S., Afat S., Lingg A., et al. Diagnostic confidence and feasibility of a deep learning accelerated haste sequence of the abdomen in a single breath-hold. Invest Radiol. 2021;56:313–319. doi: 10.1097/RLI.0000000000000743. [DOI] [PubMed] [Google Scholar]
- 23.Yu M., Tang X., Jin H., Yang S., Yun H., Wang Q., et al. Coronary ct angiography in asymptomatic adults with hepatic steatosis. Radiology. 2021;301:593–601. doi: 10.1148/radiol.2021210355. [DOI] [PubMed] [Google Scholar]
- 24.Chen Z., Xue Y., Wu Y., Duan Q., Zheng E., He Y., et al. Feasibility of 3d breath-hold mr cholangiopancreatography with a spatially selective radiofrequency excitation pulse: prospective comparison with parallel imaging technique and compressed sensing method. Acad Radiol. 2022;29:e289–e295. doi: 10.1016/j.acra.2022.03.006. [DOI] [PubMed] [Google Scholar]
- 25.Song Q., Shi Y., Gao F., Yin M., Yang R., Liu Y., et al. Feasibility and reproducibility of multifrequency magnetic resonance elastography in healthy and diseased pancreases. J Magn Reson Imaging. 2022;56:1769–1780. doi: 10.1002/jmri.28158. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Tajima T., Akai H., Yasaka K., Kunimatsu A., Akahane M., Yoshioka N., et al. Clinical feasibility of an abdominal thin-slice breath-hold single-shot fast spin echo sequence processed using a deep learning-based noise-reduction approach. Magn Reson Imaging. 2022;90:76–83. doi: 10.1016/j.mri.2022.04.005. [DOI] [PubMed] [Google Scholar]
- 27.Hong G.S., Byun J.H., Kim J.H., Kim H.J., Lee S.S., Hong S.M., et al. Thread sign in biliary intraductal papillary mucinous neoplasm: a novel specific finding for mri. Eur Radiol. 2016;26:3112–3120. doi: 10.1007/s00330-015-4158-5. [DOI] [PubMed] [Google Scholar]
- 28.Njølstad T., Schulz A., Jensen K., Andersen H.K., Martinsen A.C.T. Improved image quality with deep learning reconstruction – a study on a semi-anthropomorphic upper-abdomen phantom. ReDII. 2023;5 doi: 10.1016/j.redii.2023.100022. [DOI] [Google Scholar]
- 29.Chaika M., Afat S., Wessling D., Afat C., Nickel D., Kannengiesser S., et al. Deep learning-based super-resolution gradient echo imaging of the pancreas: improvement of image quality and reduction of acquisition time. Diagn Interv Imaging. 2023;104:53–59. doi: 10.1016/j.diii.2022.06.006. [DOI] [PubMed] [Google Scholar]
- 30.Ahmed T.M., Kawamoto S., Hruban R.H., Fishman E.K., Soyer P., Chu L.C. A primer on artificial intelligence in pancreatic imaging. Diagn Interv Imaging. 2023;104:435–447. doi: 10.1016/j.diii.2023.03.002. [DOI] [PubMed] [Google Scholar]
- 31.Aggarwal H.K., Mani M.P., Jacob M. Modl: model-based deep learning architecture for inverse problems. IEEE Trans Med Imaging. 2019;38:394–405. doi: 10.1109/TMI.2018.2865356. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Zhou Z., Han F., Ghodrati V., Gao Y., Yin W., Yang Y., et al. Parallel imaging and convolutional neural network combined fast MR image reconstruction: applications in low-latency accelerated real-time imaging. Medical Physics. 2019;46:3399–3413. doi: 10.1002/mp.13628. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Sun W., Wang W., Zhu K., Chen C., Wen X., Zeng M., et al. Feasibility of compressed sensing technique for isotropic dynamic contrast-enhanced liver magnetic resonance imaging. Eur J Radiol. 2021;139 doi: 10.1016/j.ejrad.2021.109729. [DOI] [PubMed] [Google Scholar]
- 34.Han S., Lee J.M., Kim S.W., Park S., Nickel M.D., Yoon J.H. Evaluation of haste T2 weighted image with reduced echo time for detecting focal liver lesions in patients at risk of developing hepatocellular carcinoma. Eur J Radiol. 2022;157 doi: 10.1016/j.ejrad.2022.110588. [DOI] [PubMed] [Google Scholar]
- 35.Kromrey M., Funayama S., Tamada D., Ichikawa S., Shimizu T., Onishi H., et al. Clinical evaluation of respiratory-triggered 3d mrcp with navigator echoes compared to breath-hold acquisition using compressed sensing and/or parallel imaging. Magn Reson Med Sci. 2020;19:318–323. doi: 10.2463/mrms.mp-2019-0122. [DOI] [PMC free article] [PubMed] [Google Scholar]