Deep Learning-based Reconstruction Research Articles

Deep learning reconstruction for optimized bone assessment in zero echo time MR imaging of the knee

PurposeTo evaluate the impact of deep learning-based reconstruction (DLRecon) on bone assessment in zero echo-time (ZTE) MRI of the knee at 1.5 Tesla. MethodsThis retrospective study included 48 consecutive exams of 46 patients (23 females) who underwent clinically indicated knee MRI at 1.5 Tesla. Standard imaging protocol comprised a sagittal prescribed, isotropic ZTE sequence. ZTE image reconstruction was performed with a standard-of-care (non-DL) and prototype DLRecon method. Exams were divided into subsets with and without osseous pathology based on the radiology report. Using a 4-point scale, two blinded readers qualitatively graded features of bone depiction including artifacts and conspicuity of pathology including diagnostic certainty in the respective subsets. Quantitatively, one reader measured signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of bone. Comparative analyses were conducted to assess the differences between the reconstruction methods. In addition, interreader agreement was calculated for the qualitative gradings. ResultsDLRecon significantly improved gradings for bone depiction relative to non-DL reconstruction (all, p < 0.05), while there was no significant difference with regards to artifacts (both, median score of 0; p = 0.058). In the subset with pathologies, conspicuity of pathology and diagnostic confidence were also scored significantly higher in DLRecon compared to non-DL (median 3 vs 2; p ≤ 0.03). Interreader agreement ranged from moderate to almost-perfect (κ = 0.54–0.88). Quantitatively, DLRecon demonstrated significantly enhanced CNR and SNR of bone compared to non-DL (p < 0.001). ConclusionZTE MRI with DLRecon improved bone depiction in the knee, compared to non-DL. Additionally, DLRecon increased conspicuity of osseous findings together with diagnostic certainty.

An unrolled neural network for accelerated dynamic MRI based on second-order half-quadratic splitting model

The reconstruction of dynamic magnetic resonance images from incomplete k-space data has sparked significant research interest due to its potential to reduce scan time. However, traditional iterative optimization algorithms fail to faithfully reconstruct images at higher acceleration factors and incur long reconstruction time. Furthermore, end-to-end deep learning-based reconstruction algorithms suffer from large model parameters and lack robustness in the reconstruction results. Recently, unrolled deep learning models, have shown immense potential in algorithm stability and applicability flexibility. In this paper, we propose an unrolled deep learning network based on a second-order Half-Quadratic Splitting(HQS) algorithm, where the forward propagation process of this framework strictly follows the computational flow of the HQS algorithm. In particular, we propose a degradation-sense module by associating random sampling patterns with intermediate variables to guide the iterative process. We introduce the Information Fusion Transformer(IFT) to extract both local and non-local prior information from image sequences, thereby removing aliasing artifacts resulting from random undersampling. Finally, we impose low-rank constraints within the HQS algorithm to further enhance the reconstruction results. The experiments demonstrate that each component module of our proposed model contributes to the improvement of the reconstruction task. Our proposed method achieves comparably satisfying performance to the state-of-the-art methods and it exhibits excellent generalization capabilities across different sampling masks. At the low acceleration factor, there is a 0.7% enhancement in the PSNR. Furthermore, when the acceleration factor reached 8 and 12, the PSNR achieves an improvement of 3.4% and 5.8% respectively.

Assessment of image quality and diagnostic accuracy for cervical spondylosis using T2w-STIR sequence with a deep learning-based reconstruction approach.

To investigate potential of enhancing image quality, maintaining interobserver consensus, and elevating disease diagnostic efficacy through the implementation of deep learning-based reconstruction (DLR) processing in 3.0T cervical spine fast magnetic resonance imaging (MRI) images, compared with conventional images. The 3.0T cervical spine MRI images of 71 volunteers were categorized into two groups: sagittal T2-weighted short T1 inversion recovery without DLR (Sag T2w-STIR) and with DLR (Sag T2w-STIR-DLR). The assessment covered artifacts, perceptual signal-to-noise ratio, clearness of tissue interfaces, fat suppression, overall image quality, and the delineation of spinal cord, vertebrae, discs, dopamine, and joints. Spanning canal stenosis, neural foraminal stenosis, herniated discs, annular fissures, hypertrophy of the ligamentum flavum or vertebral facet joints, and intervertebral disc degeneration were evaluated by three impartial readers. Sag T2w-STIR-DLR images exhibited markedly superior performance across quality indicators (median = 4 or 5) compared to Sag T2w-STIR sequences (median = 3 or 4) (p < 0.001). No statistically significant differences were observed between the two sequences in terms of diagnosis and grading (p > 0.05). The interobserver agreement for Sag T2w-STIR-DLR images (0.604-0.931) was higher than the other (0.545-0.853), Sag T2w-STIR-DLR (0.747-1.000) demonstrated increased concordance between reader 1 and reader 3 in comparison to Sag T2w-STIR (0.508-1.000). Acquisition time diminished from 364 to 197s through the DLR scheme. Our investigation establishes that 3.0T fast MRI images subjected to DLR processing present heightened image quality, bolstered diagnostic performance, and reduced scanning durations for cervical spine MRI compared with conventional sequences.

Deep Learning-Based Reconstruction of 3D T1 SPACE Vessel Wall Imaging Provides Improved Image Quality with Reduced Scan Times: A Preliminary Study.

Intracranial vessel wall imaging is technically challenging to implement, given the simultaneous requirements of high spatial resolution, excellent blood and CSF signal suppression, and clinically acceptable gradient times. Herein, we present our preliminary findings on the evaluation of a deep learning-optimized sequence using T1-weighted imaging. Clinical and optimized deep learning-based image reconstruction T1 3D Sampling Perfection with Application optimized Contrast using different flip angle Evolution (SPACE) were evaluated, comparing noncontrast sequences in 10 healthy controls and postcontrast sequences in 5 consecutive patients. Images were reviewed on a Likert-like scale by 4 fellowship-trained neuroradiologists. Scores (range, 1-4) were separately assigned for 11 vessel segments in terms of vessel wall and lumen delineation. Additionally, images were evaluated in terms of overall background noise, image sharpness, and homogeneous CSF signal. Segment-wise scores were compared using paired samples t tests. The scan time for the clinical and deep learning-based image reconstruction sequences were 7:26 minutes and 5:23 minutes respectively. Deep learning-based image reconstruction images showed consistently higher wall signal and lumen visualization scores, with the differences being statistically significant in most vessel segments on both pre- and postcontrast images. Deep learning-based image reconstruction had lower background noise, higher image sharpness, and uniform CSF signal. Depiction of intracranial pathologies was better or similar on the deep learning-based image reconstruction. Our preliminary findings suggest that deep learning-based image reconstruction-optimized intracranial vessel wall imaging sequences may be helpful in achieving shorter gradient times with improved vessel wall visualization and overall image quality. These improvements may help with wider adoption of intracranial vessel wall imaging in clinical practice and should be further validated on a larger cohort.

Deep learning-based low-dose CT simulator for non-linear reconstruction methods.

Computer algorithms that simulate lower-doses computed tomography (CT) images from clinical-dose images are widely available. However, most operate in the projection domain and assume access to the reconstruction method. Access to commercial reconstruction methods may often not be available in medical research, making image-domain noise simulation methods useful. However, the introduction of non-linear reconstruction methods, such as iterative and deep learning-based reconstruction, makes noise insertion in the image domain intractable, as it is not possible to determine the noise textures analytically. To develop a deep learning-based image-domain method to generate low-dose CT images from clinical-dose CT (CDCT) images for non-linear reconstruction methods. We propose a fully image domain-based method, utilizing a series of three convolutional neural networks (CNNs), which, respectively, denoise CDCT images, predict the standard deviation map of the low-dose image, and generate the noise power spectra (NPS) of local patches throughout the low-dose image. All three models have U-net-based architectures and are partly or fully three-dimensional. As a use case for this study and with no loss of generality, we use paired low-dose and clinical-dose brain CT scans. A dataset of paired scans was retrospectively obtained. All images were acquired with a wide-area detector clinical system and reconstructed using its standard clinical iterative algorithm. Each pair was registered using rigid registration to correct for motion between acquisitions. The data was randomly partitioned into training ( samples), validation ( samples), and test ( samples) sets. The performance of each of these three CNNs was validated separately. For the denoising CNN, the local standard deviation decrease, and bias were determined. For the standard deviation map CNN, the real and estimated standard deviations were compared locally. Finally, for the NPS CNN, the NPS of the synthetic and real low-dose noise were compared inside and outside the skull. Two proof-of-concept denoising studies were performed to determine if the performance of a CNN- or a gradient-based denoising filter on the synthetic low-dose data versus real data differed. The denoising network had a median decrease in noise in the cerebrospinal fluid by a factor of and introduced a median bias of HU. The network for standard deviation map estimation had a median error of HU. The noise power spectrum estimation network was able to capture the anisotropic and shift-variant nature of the noise structure by showing good agreement between the synthetic and real low-dose noise and their corresponding power spectra. The two proof of concept denoising studies showed only minimal difference in standard deviation improvement ratio between the synthetic and real low-dose CT images with the median difference between the two being 0.0 and +0.05 for the CNN- and gradient-based filter, respectively. The proposed method demonstrated good performance in generating synthetic low-dose brain CT scans without access to the projection data or to the reconstruction method. This method can generate multiple low-dose image realizations from one clinical-dose image, so it is useful for validation, optimization, and repeatability studies of image-processing algorithms.

Evaluation of deep-learning TSE images in clinical musculoskeletal imaging.

In this study, we compared the fat-saturated (FS) and non-FS turbo spin echo (TSE) magnetic resonance imaging knee sequences reconstructed conventionally (conventional-TSE) against a deep learning-based reconstruction of accelerated TSE (DL-TSE) scans. A total of 232 conventional-TSE and DL-TSE image pairs were acquired for comparison. For each consenting patient, one of the clinically acquired conventional-TSE proton density-weighted sequences in the sagittal or coronal planes (FS and non-FS), or in the axial plane (non-FS), was repeated using a research DL-TSE sequence. The DL-TSE reconstruction resulted in an image resolution that increased by at least 45% and scan times that were up to 52% faster compared to the conventional TSE. All images were acquired on a MAGNETOM Vida 3T scanner (Siemens Healthineers AG, Erlangen, Germany). The reporting radiologists, blinded to the acquisition time, were requested to qualitatively compare the DL-TSE against the conventional-TSE reconstructions. Despite having a faster acquisition time, the DL-TSE was rated to depict smaller structures better for 139/232 (60%) cases, equivalent for 72/232 (31%) cases and worse for 21/232 (9%) cases compared to the conventional-TSE. Overall, the radiologists preferred the DL-TSE reconstruction in 124/232 (53%) cases and stated no preference, implying equivalence, for 65/232 (28%) cases. DL-TSE reconstructions enabled faster acquisition times while enhancing spatial resolution and preserving the image contrast. From these results, the DL-TSE provided added or comparable clinical value and utility in less time. DL-TSE offers the opportunity to further reduce the overall examination time and improve patient comfort with no loss in diagnostic accuracy.

Evaluation of deep learning-based reconstruction late gadolinium enhancement images for identifying patients with clinically unrecognized myocardial infarction

BackgroundThe presence of infarction in patients with unrecognized myocardial infarction (UMI) is a critical feature in predicting adverse cardiac events. This study aimed to compare the detection rate of UMI using conventional and deep learning reconstruction (DLR)-based late gadolinium enhancement (LGEO and LGEDL, respectively) and evaluate optimal quantification parameters to enhance diagnosis and management of suspected patients with UMI.MethodsThis prospective study included 98 patients (68 men; mean age: 55.8 ± 8.1 years) with suspected UMI treated at our hospital from April 2022 to August 2023. LGEO and LGEDL images were obtained using conventional and commercially available inline DLR algorithms. The myocardial signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and percentage of enhanced area (Parea) employing the signal threshold versus reference mean (STRM) approach, which correlates the signal intensity (SI) within areas of interest with the average SI of normal regions, were analyzed. Analysis was performed using the standard deviation (SD) threshold approach (2SD–5SD) and full width at half maximum (FWHM) method. The diagnostic efficacies based on LGEDL and LGEO images were calculated.ResultsThe SNRDL and CNRDL were two times better than the SNRO and CNRO, respectively (P < 0.05). Parea−DL was elevated compared to Parea−O using the threshold methods (P < 0.05); however, no intergroup difference was found based on the FWHM method (P > 0.05). The Parea−DL and Parea−O also differed except between the 2SD and 3SD and the 4SD/5SD and FWHM methods (P < 0.05). The receiver operating characteristic curve analysis revealed that each SD method exhibited good diagnostic efficacy for detecting UMI, with the Parea−DL having the best diagnostic efficacy based on the 5SD method (P < 0.05). Overall, the LGEDL images had better image quality. Strong diagnostic efficacy for UMI identification was achieved when the STRM was ≥ 4SD and ≥ 3SD for the LGEDL and LGEO, respectively.ConclusionsSTRM selection for LGEDL magnetic resonance images helps improve clinical decision-making in patients with UMI. This study underscored the importance of STRM selection for analyzing LGEDL images to enhance diagnostic accuracy and clinical decision-making for patients with UMI, further providing better cardiovascular care.

Deep learning-based spectroscopic single-molecule localization microscopy.

Spectroscopic single-molecule localization microscopy (sSMLM) takes advantage of nanoscopy and spectroscopy, enabling sub-10nm resolution as well as simultaneous multicolor imaging of multi-labeled samples. Reconstruction of raw sSMLM data using deep learning is a promising approach for visualizing the subcellular structures at the nanoscale. Develop a novel computational approach leveraging deep learning to reconstruct both label-free and fluorescence-labeled sSMLM imaging data. We developed a two-network-model based deep learning algorithm, termed DsSMLM, to reconstruct sSMLM data. The effectiveness of DsSMLM was assessed by conducting imaging experiments on diverse samples, including label-free single-stranded DNA (ssDNA) fiber, fluorescence-labeled histone markers on COS-7 and U2OS cells, and simultaneous multicolor imaging of synthetic DNA origami nanoruler. For label-free imaging, a spatial resolution of 6.22nm was achieved on ssDNA fiber; for fluorescence-labeled imaging, DsSMLM revealed the distribution of chromatin-rich and chromatin-poor regions defined by histone markers on the cell nucleus and also offered simultaneous multicolor imaging of nanoruler samples, distinguishing two dyes labeled in three emitting points with a separation distance of 40nm. With DsSMLM, we observed enhanced spectral profiles with 8.8% higher localization detection for single-color imaging and up to 5.05% higher localization detection for simultaneous two-color imaging. We demonstrate the feasibility of deep learning-based reconstruction for sSMLM imaging applicable to label-free and fluorescence-labeled sSMLM imaging data. We anticipate our technique will be a valuable tool for high-quality super-resolution imaging for a deeper understanding of DNA molecules' photophysics and will facilitate the investigation of multiple nanoscopic cellular structures and their interactions.

Effectiveness of deep learning-based reconstruction for improvement of image quality and liver tumor detectability in the hepatobiliary phase of gadoxetic acid-enhanced magnetic resonance imaging.

To evaluate the effectiveness of deep learning-based reconstruction (DLR) in improving image quality and tumor detectability of isovoxel high-resolution breath-hold fat-suppressed T1-weighted imaging (HR-BH-FS-T1WI) in the hepatobiliary phase (HBP) of Gadoxetic acid-enhanced magnetic resonance imaging (Gd-EOB-MRI). This retrospective evaluated 42 patients with 98 liver tumors who underwent Gd-EOB-MRI between March 2023 and May 2023 using three techniques based on HBP imaging: isovoxel HR-BH-FS-T1WI reconstructed (1) with DLR (BH-DLR +) and (2) without DLR (BH-DLR -) and (3) HR-FS-T1WI scanned with a free-breathing technique using a navigator-echo-triggered technique and DLR (Navi-DLR +). The three techniques were qualitatively and quantitatively compared by the Friedman test and the Bonferroni post-hoc test. Tumor detectability was compared using the McNemar test. BH-DLR + (3.85, average score of two radiologists) showed significantly better qualitative scores for image noise than BH-DLR - (2.84) and Navi-DLR + (3.37) (p < 0.0167), and Navi-DLR + showed significantly better scores than BH-DLR - (p < 0.0167). BH-DLR + (3.77) and BH-DLR - (3.77) showed significantly better qualitative scores for respiratory motion artifact than Navi-DLR + (2.75) (p < 0.0167), but there was no significant difference in scores between BH-DLR + and BH-DLR - (p > 0.0167). BH-DLR + (0.32) and Navi-DLR + (0.33) showed significantly higher lesion-to-nonlesion CR than BH-DLR - (0.29) (p < 0.0167), but there was no significant difference in lesion-to-nonlesion CR between BH-DLR + and Navi-DLR + (p > 0.0167). BH-DLR + (89.8%) showed significantly better tumor detectability than BH-DLR - (76.0%) and Navi-DLR + (77.6%) (p < 0.05). The use of DLR for isovoxel HR-BH-FS-T1WI was effective in improving image quality and tumor detectability.

Improved stent sharpness evaluation with super-resolution deep learning reconstruction in coronary CT angiography.

This study aimed to assess the impact of super-resolution deep learning reconstruction (SR-DLR) on coronary CT angiography (CCTA) image quality and blooming artifacts from coronary artery stents in comparison to conventional methods, including hybrid iterative reconstruction (HIR) and deep learning-based reconstruction (DLR). A retrospective analysis included 66 CCTA patients from July to November 2022. Major coronary arteries were evaluated for image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). Stent sharpness was quantified using 10%-90% edge rise slope (ERS) and 10%-90% edge rise distance (ERD). Qualitative analysis employed a 5-point scoring system to assess overall image quality, image noise, vessel wall, and stent structure. SR-DLR demonstrated significantly lower image noise compared to HIR and DLR. SNR and CNR were notably higher in SR-DLR. Stent ERS was significantly improved in SR-DLR, with mean ERD values of 0.70 ± 0.20 mm for SR-DLR, 1.13 ± 0.28 mm for HIR, and 0.85 ± 0.26 mm for DLR. Qualitatively, SR-DLR scored higher in all categories. SR-DLR produces images with lower image noise, leading to improved overall image quality, compared with HIR and DLR. SR-DLR is a valuable image reconstruction algorithm for enhancing the spatial resolution and sharpness of coronary artery stents without being constrained by hardware limitations. The overall image quality was significantly higher in SR-DLR, resulting in sharper coronary artery stents compared to HIR and DLR.

Deep-learning-based reconstruction of T2-weighted magnetic resonance imaging of the prostate accelerated by compressed sensing provides improved image quality at half the acquisition time.

Deep-learning-based reconstruction (DLR) improves the quality of magnetic resonance (MR) images which allows faster acquisitions. The aim of this study was to compare the image quality of standard and accelerated T2 weighted turbo-spin-echo (TSE) images of the prostate reconstructed with and without DLR and to find associations between perceived image quality and calculated image characteristics. In a cohort of 47 prospectively enrolled consecutive patients referred for bi-parametric prostate magnetic resonance imaging (MRI), two T2-TSE acquisitions in the transverse plane were acquired on a 3T scanner-a standard T2-TSE sequence and a short sequence accelerated by a factor of two using compressed sensing (CS). The images were reconstructed with and without DLR in super-resolution mode. The image quality was rated in six domains. Signal-to-noise ratio (SNR), and image sharpness were measured. The mean acquisition time was 281±23 s for the standard and 140±12 s for the short acquisition (P<0.0001). DLR images had higher sharpness compared to non-DLR (P<0.001). Short and short-DLR had lower SNR than the standard and standard-DLR (P<0.001). The perceived image quality of short-DLR was rated better in all categories compared to the standard sequence (P<0.001 to P=0.004). All domains of subjective evaluation were correlated with measured image sharpness (P<0.001). T2-TSE acquisition of the prostate accelerated using CS combined with DLR reconstruction provides images with increased sharpness that have a superior quality as perceived by human readers compared to standard T2-TSE. The perceived image quality is correlated with measured image contrast.

Radiation dose reduction and image quality improvement with ultra-high resolution temporal bone CT using deep learning-based reconstruction: An anatomical study

PurposeThe purpose of this study was to evaluate the achievable radiation dose reduction of an ultra-high resolution computed tomography (UHR-CT) scanner using deep learning reconstruction (DLR) while maintaining temporal bone image quality equal to or better than high-resolution CT (HR-CT). Materials and methodsUHR-CT acquisitions were performed with variable tube voltages and currents at eight different dose levels (volumic CT dose index [CTDIvol] range: 4.6–79 mGy), 10242 matrix, and 0.25 mm slice thickness and reconstructed using DLR and hybrid iterative reconstruction (HIR) algorithms. HR-CT images were acquired using a standard protocol (120 kV/220 mAs; CTDI vol, 54.2 mGy, 5122 matrix, and 0.5 mm slice thickness). Two radiologists rated the image quality of seven structures using a five point confidence scale on six cadaveric temporal bone CTs. A global image quality score was obtained for each CT protocol by summing the image quality scores of all structures. ResultsWith DLR, UHR-CT at 120 kV/220 mAs (CTDIvol, 50.9 mGy) and 140 kV/220 mAs (CTDIvol, 79 mGy) received the highest global image quality scores (4.88 ± 0.32 [standard deviation (SD)] [range: 4–5] and 4.85 ± 0.35 [range: 4–5], respectively; P = 0.31), while HR-CT at 120 kV/220 mAs and UHR-CT at 120 kV/20 mAs received the lowest (i.e., 3.14 ± 0.75 [SD] [range: 2–5] and 2.97 ± 0.86 [SD] [range: 1–5], respectively; P = 0.14). All the DLR protocols had better image quality scores than HR-CT with HIR. ConclusionUHR-CT with DLR can be performed with up to a tenfold reduction in radiation dose compared to HR-CT with HIR while maintaining or improving image quality.

Advanced deep learning-based image reconstruction in lumbar spine MRI at 0.55 T – Effects on image quality and acquisition time in comparison to conventional deep learning-based reconstruction

ObjectivesTo evaluate an optimized deep leaning-based image post-processing technique in lumbar spine MRI at 0.55 T in terms of image quality and image acquisition time. Materials and methodsLumbar spine imaging was conducted on 18 patients using a 0.55 T MRI scanner, employing conventional (CDLR) and advanced (ADLR) deep learning-based post-processing techniques. Two musculoskeletal radiologists visually evaluated the images using a 5-point Likert scale to assess image quality and resolution. Quantitative assessment in terms of signal intensities (SI) and contrast ratios was performed by region of interest measurements in different body-tissues (vertebral bone, intervertebral disc, spinal cord, cerebrospinal fluid and autochthonous back muscles) to investigate differences between CDLR and ADLR sequences. ResultsThe images processed with the advanced technique (ADLR) were rated superior to the conventional technique (CDLR) in terms of signal/contrast, resolution, and assessability of the spinal canal and neural foramen. The interrater agreement was moderate for signal/contrast (ICC = 0.68) and good for resolution (ICC = 0.77), but moderate for spinal canal and neuroforaminal assessability (ICC = 0.55). Quantitative assessment showed a higher contrast ratio for fluid-sensitive sequences in the ADLR images. The use of ADLR reduced image acquisition time by 44.4%, from 14:22 min to 07:59 min. ConclusionsAdvanced deep learning-based image reconstruction algorithms improve the visually perceived image quality in lumbar spine imaging at 0.55 T while simultaneously allowing to substantially decrease image acquisition times. Clinical relevanceAdvanced deep learning-based image post-processing techniques (ADLR) in lumbar spine MRI at 0.55 T significantly improves image quality while reducing image acquisition time.

Enhanced bone assessment of the shoulder using zero-echo time MRI with deep-learning image reconstruction

ObjectivesTo assess a deep learning-based reconstruction algorithm (DLRecon) in zero echo-time (ZTE) MRI of the shoulder at 1.5 Tesla for improved delineation of osseous findings.MethodsIn this retrospective study, 63 consecutive exams of 52 patients (28 female) undergoing shoulder MRI at 1.5 Tesla in clinical routine were included. Coronal 3D isotropic radial ZTE pulse sequences were acquired in the standard MR shoulder protocol. In addition to standard-of-care (SOC) image reconstruction, the same raw data was reconstructed with a vendor-supplied prototype DLRecon algorithm. Exams were classified into three subgroups: no pathological findings, degenerative changes, and posttraumatic changes, respectively. Two blinded readers performed bone assessment on a 4-point scale (0-poor, 3-perfect) by qualitatively grading image quality features and delineation of osseous pathologies including diagnostic confidence in the respective subgroups. Quantitatively, signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of bone were measured. Qualitative variables were compared using the Wilcoxon signed‐rank test for ordinal data and the McNemar test for dichotomous variables; quantitative measures were compared with Student’s t-testing.ResultsDLRecon scored significantly higher than SOC in all visual metrics of image quality (all, p < 0.03), except in the artifact category (p = 0.37). DLRecon also received superior qualitative scores for delineation of osseous pathologies and diagnostic confidence (p ≤ 0.03). Quantitatively, DLRecon achieved superior CNR (95 CI [1.4–3.1]) and SNR (95 CI [15.3–21.5]) of bone than SOC (p < 0.001).ConclusionDLRecon enhanced image quality in ZTE MRI and improved delineation of osseous pathologies, allowing for increased diagnostic confidence in bone assessment.

Bladder MRI with deep learning-based reconstruction: a prospective evaluation of muscle invasiveness using VI-RADS.

To investigate the influence of deep learning reconstruction (DLR) on bladder MRI, specifically examination time, image quality, and diagnostic performance of vesical imaging reporting and data system (VI-RADS) within a prospective clinical cohort. Seventy participants with bladder cancer who underwent MRI between August 2022 and February 2023 with a protocol containing standard T2-weighted imaging (T2WIS), standard diffusion-weighted imaging (DWIS), fast T2WI with DLR (T2WIDL), and fast DWI with DLR (DWIDL) were enrolled in this prospective study. Imaging quality was evaluated by measuring signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and qualitative image quality scoring. Additionally, the apparent diffusion coefficient (ADC) of bladder lesions derived from DWIS and DWIDL was measured and VI-RADS scoring was performed. Paired t-test or paired Wilcoxon signed-rank test were performed to compare image quality score, SNR, CNR, and ADC between standard sequences and fast sequences with DLR. The diagnostic performance for VI-RADS was assessed using the areaunder the receiver operating characteristiccurve(AUC). Compared to T2WIS and DWIS, T2WIDL and DWIDL reduced the acquisition time from 5:57min to 3:13min and showed significantly higher SNR, CNR, qualitative image quality score of overall image quality, image sharpness, and lesion conspicuity. There were no significant differences in ADC and AUC of VI-RADS between standard sequences and fast sequences with DLR. The application of DLR to T2WI and DWI reduced examination time and significantly improved image quality, maintaining ADC and the diagnostic performance of VI-RADS for evaluating muscle invasion in bladder cancer.

Reduction of ADC bias in diffusion MRI with deep learning-based acceleration: A phantom validation study at 3.0 T

PurposeFurther acceleration of DWI in diagnostic radiology is desired but challenging mainly due to low SNR in high b-value images and associated bias in quantitative ADC values. Deep learning-based reconstruction and denoising may provide a solution to address this challenge. MethodsThe effects of SNR reduction on ADC bias and variability were investigated using a commercial diffusion phantom and numerical simulations. In the phantom, performance of different reconstruction methods, including conventional parallel (SENSE) imaging, compressed sensing (C-SENSE), and compressed SENSE acceleration with an artificial intelligence deep learning-based technique (C-SENSE AI), was compared at different acceleration factors and flip angles using ROI-based analysis. ADC bias was assessed by Lin's Concordance correlation coefficient (CCC) followed by bootstrapping to calculate confidence intervals (CI). ADC random measurement error (RME) was assessed by the mean coefficient of variation (CV¯) and non-parametric statistical tests. ResultsThe simulations predicted increasingly negative bias and loss of precision towards lower SNR. These effects were confirmed in phantom measurements of increasing acceleration, for which CCC decreased from 0.947 to 0.279 and CV¯ increased from 0.043 to 0.439, and of decreasing flip angle, for which CCC decreased from 0.990 to 0.063 and CV¯ increased from 0.037 to 0.508. At high acceleration and low flip angle, C-SENSE AI reconstruction yielded best denoised ADC maps. For the lowest investigated flip angle, CCC = {0.630, 0.771 and 0.987} and CV¯={0.508, 0.426 and 0.254} were obtained for {SENSE, C-SENSE, C-SENSE AI}, the improvement by C-SENSE AI being significant as compared to the other methods (CV: p = 0.033 for C-SENSE AI vs. C-SENSE and p < 0.001 for C-SENSE AI vs. SENSE; CCC: non-overlapping CI between reconstruction methods). For the highest investigated acceleration factor, CCC = {0.479,0.926,0.960} and CV¯={0.519,0.119,0.118} were found, confirming the reduction of bias and RME by C-SENSE AI as compared to C-SENSE (by trend) and to SENSE (CV: p < 0.001; CCC: non-overlapping CI). ConclusionADC bias and random measurement error in DWI at low SNR, typically associated with scan acceleration, can be effectively reduced by deep-learning based C-SENSE AI reconstruction.

Evaluation of four computed tomography reconstruction algorithms using a coronary artery phantom.

Despite advancements in coronary computed tomography angiography (CTA), challenges in positive predictive value and specificity remain due to limited spatial resolution. The purpose of this experimental study was to investigate the effect of 2nd generation deep learning-based reconstruction (DLR) on the quantitative and qualitative image quality in coronary CTA. A vessel model with stepwise non-calcified plaque was scanned using 320-detector CT. Image reconstruction was performed using four techniques: hybrid iterative reconstruction (HIR), model-based iterative reconstruction (MBIR), DLR, and 2nd generation DLR. The luminal peak CT number, contrast-to-noise ratio (CNR), and edge rise slope (ERS) were quantitatively evaluated via profile curve analysis. Two observers qualitatively graded the graininess, lumen sharpness, and overall lumen visibility on the basis of the degree of confidence for the stenosis severity using a five-point scale. The image noise with HIR, MBIR, DLR, and 2nd generation DLR was 23.0, 21.0, 16.9, and 9.5 HU, respectively. The corresponding CNR (25% stenosis) was 15.5, 15.9, 22.1, and 38.3, respectively. The corresponding ERS (25% stenosis) was 203.2, 198.6, 228.9, and 262.4 HU/mm, respectively. Among the four reconstruction methods, the 2nd generation DLR achieved the significantly highest CNR and ERS values. The score of 2nd generation DLR in all evaluation points (graininess, sharpness, and overall lumen visibility) was higher than those of the other methods (overall vessel visibility score, 2.6±0.5, 3.8±0.6, 3.7±0.5, and 4.6±0.5 with HIR, MBIR, DLR, and 2nd generation DLR, respectively). 2nd generation DLR provided better CNR and ERS in coronary CTA than HIR, MBIR, and previous-generation DLR, leading to the highest subjective image quality in the assessment of vessel stenosis.

Deep learning-based reconstruction enhances image quality and improves diagnosis in magnetic resonance imaging of the shoulder joint.

Accelerated magnetic resonance imaging sequences reconstructed using the vendor-provided Recon deep learning algorithm (DL-MRI) were found to be more likely than conventional magnetic resonance imaging (MRI) sequences to detect subacromial (SbA) bursal thickening. However, the extent of this thickening was not extensively explored. This study aimed to compare the image quality of DL-MRI with conventional MRI sequences reconstructed via conventional pipelines (Conventional-MRI) for shoulder examinations and evaluate the effectiveness of DL-MRI in accurately demonstrating the degree of SbA bursal and subcoracoid (SC) bursal thickening. This prospective cross-sectional study enrolled 41 patients with chronic shoulder pain who underwent 3-T MRI (including both Conventional-MRI and accelerated MRI sequences) of the shoulder between December 2022 and April 2023. Each protocol consisted of oblique axial, coronal, and sagittal images, including proton density-weighted imaging (PDWI) with fat suppression (FS) and oblique coronal T1-weighted imaging (T1WI) with FS. The image quality and degree of artifacts were assessed using a 5-point Likert scale for both Conventional-MRI and DL-MRI. Additionally, the degree of SbA and SC bursal thickening, as well as the relative signal-to-noise ratio (rSNR) and relative contrast-to-noise ratio (rCNR) were analyzed using the paired Wilcoxon test. Statistical significance was defined as P<0.05. The utilization of accelerated sequences resulted in a remarkable 54.7% reduction in total scan time. Overall, DL-MRI exhibited superior image quality scores and fewer artifacts compared to Conventional-MRI. Specifically, DL-MRI demonstrated significantly higher measurements of SC bursal thickenings in the oblique sagittal PDWI sequence compared to Conventional-MRI [3.92 (2.83, 5.82) vs. 3.74 (2.92, 5.96) mm, P=0.028]. Moreover, DL-MRI exhibited higher detection of SbA bursal thickenings in the oblique coronal PDWI sequence [2.61 (1.85, 3.46) vs. 2.48 (1.84, 3.25) mm], with a notable trend towards significant differences (P=0.071). Furthermore, the rSNRs of the muscle in DL-MRI images were significantly higher than those in Conventional-MRI images across most sequences (P<0.001). However, the rSNRs of bone on Conventional-MRI of oblique axial and oblique coronal PDWI sequences showed adverse results [oblique axial: 1.000 (1.000, 1.000) vs. 0.444 (0.367, 0.523); and oblique coronal: 1.000 (1.000, 1.000) vs. 0.460 (0.387, 0.631); all P<0.001]. Additionally, all DL-MRI images exhibited significantly greater rSNRs and rCNRs compared to accelerated MRI sequences reconstructed using traditional pipelines (P<0.001). In conclusion, the utilization of DL-MRI enhances image quality and improves diagnostic capabilities, making it a promising alternative to Conventional-MRI for shoulder imaging.

Highly robust reconstruction framework for three-dimensional optical imaging based on physical model constrained neural networks

Objective. The reconstruction of three-dimensional optical imaging that can quantitatively acquire the target distribution from surface measurements is a serious ill-posed problem. Traditional regularization-based reconstruction can solve such ill-posed problem to a certain extent, but its accuracy is highly dependent on a prior information, resulting in a less stable and adaptable method. Data-driven deep learning-based reconstruction avoids the errors of light propagation models and the reliance on experience and a prior by learning the mapping relationship between the surface light distribution and the target directly from the dataset. However, the acquisition of the training dataset and the training of the network itself are time consuming, and the high dependence of the network performance on the training dataset results in a low generalization ability. The objective of this work is to develop a highly robust reconstruction framework to solve the existing problems. Approach. This paper proposes a physical model constrained neural networks-based reconstruction framework. In the framework, the neural networks are to generate a target distribution from surface measurements, while the physical model is used to calculate the surface light distribution based on this target distribution. The mean square error between the calculated surface light distribution and the surface measurements is then used as a loss function to optimize the neural network. To further reduce the dependence on a priori information, a movable region is randomly selected and then traverses the entire solution interval. We reconstruct the target distribution in this movable region and the results are used as the basis for its next movement. Main Results. The performance of the proposed framework is evaluated with a series of simulations and in vivo experiment, including accuracy robustness of different target distributions, noise immunity, depth robustness, and spatial resolution. The results collectively demonstrate that the framework can reconstruct targets with a high accuracy, stability and versatility. Significance. The proposed framework has high accuracy and robustness, as well as good generalizability. Compared with traditional regularization-based reconstruction methods, it eliminates the need to manually delineate feasible regions and adjust regularization parameters. Compared with emerging deep learning assisted methods, it does not require any training dataset, thus saving a lot of time and resources and solving the problem of poor generalization and robustness of deep learning methods. Thus, the framework opens up a new perspective for the reconstruction of three-dimension optical imaging.

Exploring the Low-Dose Limit for Focal Hepatic Lesion Detection with a Deep Learning-Based CT Reconstruction Algorithm: A Simulation Study on Patient Images.

This study aims to investigate the maximum achievable dose reduction for applying a new deep learning-based reconstruction algorithm, namely the artificial intelligence iterative reconstruction (AIIR), in computed tomography (CT) for hepatic lesion detection. A total of 40 patients with 98 clinically confirmed hepatic lesions were retrospectively included. The mean volume CT dose index was 13.66 ± 1.73 mGy in routine-dose portal venous CT examinations, where the images were originally obtained with hybrid iterative reconstruction (HIR). Low-dose simulations were performed in projection domain for 40%-, 20%-, and 10%-dose levels, followed by reconstruction using both HIR and AIIR. Two radiologists were asked to detect hepatic lesion on each set of low-dose image in separate sessions. Qualitative metrics including lesion conspicuity, diagnostic confidence, and overall image quality were evaluated using a 5-point scale. The contrast-to-noise ratio (CNR) for lesion was also calculated for quantitative assessment. The lesion CNR on AIIR at reduced doses were significantly higher than that on routine-dose HIR (all p < 0.05). Lower qualitative image quality was observed as the radiation dose reduced, while there were no significant differences between 40%-dose AIIR and routine-dose HIR images. The lesion detection rate was 100%, 98% (96/98), and 73.5% (72/98) on 40%-, 20%-, and 10%-dose AIIR, respectively, whereas it was 98% (96/98), 73.5% (72/98), and 40% (39/98) on the corresponding low-dose HIR, respectively. AIIR outperformed HIR in simulated low-dose CT examinations of the liver. The use of AIIR allows up to 60% dose reduction for lesion detection while maintaining comparable image quality to routine-dose HIR.