Deep Learning-based Reconstruction Algorithms Research Articles

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-aware 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.

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

To 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. In 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. DLRecon 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). DLRecon enhanced image quality in ZTE MRI and improved delineation of osseous pathologies, allowing for increased diagnostic confidence in bone assessment.

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.

Compressed sensing framework for BCG signals based on the optical fiber sensor.

A compressed sensing (CS) framework is built for ballistocardiography (BCG) signals, which contains two parts of an optical fiber sensor-based heart monitoring system with a CS module and an end-to-end deep learning-based reconstruction algorithm. The heart monitoring system collects BCG data, and then compresses and transmits the data through the CS module at the sensing end. The deep learning-based algorithm reconstructs compressed data at the received end. To evaluate results, three traditional CS reconstruction algorithms and a deep learning method are adopted as references to reconstruct the compressed BCG data with different compression ratios (CRs). Results show that our framework can reconstruct signals successfully when the CR grows from 50% to 95% and outperforms other methods at high CRs. The mean absolute error (MAE) of the estimated heartbeat rate (HR) is lower than 1 bpm when the CR is below 95%. The proposed CS framework for BCG signals can be integrated into the IoMT system, which has great potential in health care for both medical and home use.

Image Quality and Lesion Detectability of Pancreatic Phase Thin-Slice Computed Tomography Images With a Deep Learning-Based Reconstruction Algorithm.

To evaluate the image quality and lesion detectability of pancreatic phase thin-slice computed tomography (CT) images reconstructed with a deep learning-based reconstruction (DLR) algorithm compared with filtered-back projection (FBP) and hybrid iterative reconstruction (IR) algorithms. Fifty-three patients who underwent dynamic contrast-enhanced CT including pancreatic phase were enrolled in this retrospective study. Pancreatic phase thin-slice (0.625 mm) images were reconstructed with each FBP, hybrid IR, and DLR. Objective image quality and signal-to-noise ratio of the pancreatic parenchyma, and contrast-to-noise ratio of pancreatic lesions were compared between the 3 reconstruction algorithms. Two radiologists independently assessed the image quality of all images. The diagnostic performance for the detection of pancreatic lesions was compared among the reconstruction algorithms using jackknife alternative free-response receiver operating characteristic analysis. Deep learning-based reconstruction resulted in significantly lower image noise and higher signal-to-noise ratio and contrast-to-noise ratio than hybrid IR and FBP ( P < 0.001). Deep learning-based reconstruction also yielded significantly higher visual scores than hybrid IR and FBP ( P < 0.01). The diagnostic performance of DLR for detecting pancreatic lesions was highest for both readers, although a significant difference was found only between DLR and FBP in one reader ( P = 0.02). Deep learning-based reconstruction showed improved objective and subjective image quality of pancreatic phase thin-slice CT relative to other reconstruction algorithms and has potential for improving lesion detectability.

Fast high-quality MRI protocol of the lumbar spine with deep learning-based algorithm: an image quality and scanning time comparison with standard protocol

ObjectiveThe objective of this study is to prospectively compare quantitative and subjective image quality, scanning time, and diagnostic confidence between a new deep learning-based reconstruction(DLR) algorithm and standard MRI protocol of lumbar spine.Materials and methodsEighty healthy volunteers underwent 1.5T MRI examination of lumbar spine from September 2021 to May 2023. Protocol acquisition comprised sagittal T1- and T2-weighted fast spin echo and short-tau inversion recovery images and axial multislices T2-weighted fast spin echo images. All sequences were acquired with both DLR algorithm and standard protocols. Two radiologists, blinded to the reconstruction technique, performed quantitative and qualitative image quality analysis in consensus reading; diagnostic confidence was also assessed. Quantitative image quality analysis was assessed by calculating signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). Qualitative image quality analysis and diagnostic confidence were assessed with a five-point Likert scale. Scanning times were also compared.ResultsDLR SNR was higher in all sequences (all p<0.001). CNR of the DLR was superior to conventional dataset only for axial and sagittal T2-weighted fast spin echo images (p<0.001). Qualitative analysis showed DLR had higher overall quality in all sequences (all p<0.001), with an inter-rater agreement of 0.83 (0.78–0.86).DLR total protocol scanning time was lower compared to standard protocol (6:26 vs 12:59 min, p<0.001).Diagnostic confidence for DLR algorithm was not inferior to standard protocol.ConclusionDLR applied to 1.5T MRI is a feasible method for lumbar spine imaging providing morphologic sequences with higher image quality and similar diagnostic confidence compared with standard protocol, enabling a remarkable time saving (up to 50%).

Image quality comparison of lower extremity CTA between CT routine reconstruction algorithms and deep learning reconstruction

BackgroundTo evaluate the image quality of lower extremity computed tomography angiography (CTA) with deep learning–based reconstruction (DLR) compared to model-based iterative reconstruction (MBIR), hybrid-iterative reconstruction (HIR), and filtered back projection (FBP).MethodsFifty patients (38 males, average age 59.8 ± 19.2 years) who underwent lower extremity CTA between January and May 2021 were included. Images were reconstructed with DLR, MBIR, HIR, and FBP. The standard deviation (SD), contrast-to-noise ratio (CNR), signal-to-noise ratio (SNR), noise power spectrum (NPS) curves, and the blur effect, were calculated. The subjective image quality was independently evaluated by two radiologists. The diagnostic accuracy of DLR, MBIR, HIR, and FBP reconstruction algorithms was calculated.ResultsThe CNR and SNR were significantly higher in DLR images than in the other three reconstruction algorithms, and the SD was significantly lower in DLR images of the soft tissues. The noise magnitude was the lowest with DLR. The NPS average spatial frequency (fav) values were higher using DLR than HIR. For blur effect evaluation, DLR and FBP were similar for soft tissues and the popliteal artery, which was better than HIR and worse than MBIR. In the aorta and femoral arteries, the blur effect of DLR was worse than MBIR and FBP and better than HIR. The subjective image quality score of DLR was the highest. The sensitivity and specificity of the lower extremity CTA with DLR were the highest in the four reconstruction algorithms with 98.4% and 97.2%, respectively.ConclusionsCompared to the other three reconstruction algorithms, DLR showed better objective and subjective image quality. The blur effect of the DLR was better than that of the HIR. The diagnostic accuracy of lower extremity CTA with DLR was the best among the four reconstruction algorithms.

Impact of a new deep-learning-based reconstruction algorithm on image quality in ultra-high-resolution CT: clinical observational and phantom studies.

To demonstrate the effect of an improved deep learning-based reconstruction (DLR) algorithm on Ultra-High-Resolution Computed Tomography (U-HRCT) scanners. Clinical and phantom studies were conducted. Thirty patients who underwent contrast-enhanced CT examination during the follow-up period were enrolled. Images were reconstructed using improved DLR [termed, New DLR, i.e., Advanced Intelligent Clear-IQ Engine (AiCE) Body Sharp] and conventional DLR (Conv DLR, AiCE Body) algorithms. Two radiologists assessed the overall image quality using a 5-point scale (5 = excellent; 1 = unacceptable). The noise power spectra (NPSs) were calculated to assess the frequency characteristics of the image noise, and the square root of area under the curve (√AUC NPS) between 0.05 and 0.50 cycle/mm was calculated as an indicator of the image noise. Dunnett's test was used for statistical analysis of the visual evaluation score, with statistical significance set at p < 0.05. The overall image quality of New DLR was better than that of the Conv DLR (4.2 ± 0.4 and 3.3 ± 0.4, respectively; p < 0.0001). All New DLR images had an overall image quality score above the average or excellent. The √AUCNPS value of New DLR was lower than that of Conv DLR (13.8 and 14.2, respectively). The median values of reconstruction time required with New DLR and Conv DLR were 5.0 and 7.8 min, respectively. The new DLR algorithm improved the image quality within a practical reconstruction time. The new DLR enables us to choose whether to improve image quality or reduce the dose.

Pulmonary nodule volumetric accuracy of a deep learning-based reconstruction algorithm in low-dose computed tomography: A phantom study

PurposeTo compare the image properties and pulmonary nodule volumetric accuracies among deep learning-based reconstruction (DLR), filtered back projection (FBP), and hybrid iterative reconstruction (hybrid IR) in low-dose computed tomography (LDCT). MethodsA multipurpose chest phantom containing artificial spherical pulmonary nodules with 5-, 8-, 10-, and 12-mm diameters and Hounsfield units (HUs) of −630 and +100 HU was scanned 20 times at a standard dose, based on a low-dose screening CT trial, and at 1/2, 1/6, and 1/12 of the standard dose. To assess noise reduction performance and volumetric accuracy, the standard deviations (SDs) of the pixel values and volumetric percentage errors (PEs) were compared among FBP, hybrid IR, and DLR. The noise non-stationarity index (NNSI) was calculated from 20 image replicates and compared among FBP, hybrid IR, and DLR to evaluate noise stationarity. ResultsThe SD reduction rates for FBP in hybrid IR and DLR were 62 %–85 % and 79 %–90 %, respectively. For the four nodules with +100 HU, the PE of all reconstruction methods was <±25 % (not clinically relevant). For the four nodules with −630 HU, the PEs were equivalent or lower for hybrid IR and DLR than for FBP, and the PE difference between hybrid IR and DLR ranged from 0 % to 7%. The NNSI was significantly higher for DLR than for FBP and hybrid IR (p < 0.01). ConclusionsGreater noise suppression was achieved with DLR than with hybrid IR without compromising nodule volumetric accuracy in LDCT despite the representative noise non-stationarity.

Deep Learning-Based Photoacoustic Imaging of Vascular Network Through Thick Porous Media.

Photoacoustic imaging is a promising approach used to realize in vivo transcranial cerebral vascular imaging. However, the strong attenuation and distortion of the photoacoustic wave caused by the thick porous skull greatly affect the imaging quality. In this study, we developed a convolutional neural network based on U-Net to extract the effective photoacoustic information hidden in the speckle patterns obtained from vascular network images datasets under porous media. Our simulation and experimental results show that the proposed neural network can learn the mapping relationship between the speckle pattern and the target, and extract the photoacoustic signals of the vessels submerged in noise to reconstruct high-quality images of the vessels with a sharp outline and a clean background. Compared with the traditional photoacoustic reconstruction methods, the proposed deep learning-based reconstruction algorithm has a better performance with a lower mean absolute error, higher structural similarity, and higher peak signal-to-noise ratio of reconstructed images. In conclusion, the proposed neural network can effectively extract valid information from highly blurred speckle patterns for the rapid reconstruction of target images, which offers promising applications in transcranial photoacoustic imaging.

DVS-Net: Dual-domain Variable Splitting Network for Accelerated Parallel MRI Data.

Parallel imaging is an important method to accel-erate the acquisition of magnetic resonance imaging data, which can shorten the breath-hold times and reduce motion artifacts. In this paper, we propose a joint frequency domain and image domain (dual-domain) reconstruction method by introducing the full sampling condition for the undersampled multi-coil MR data. The motivation is that the dual domain method can provide more information for accurate image reconstruction. An efficient iterative algorithm is developed based on the variable splitting technique and alternating direction method of multipliers, which is unrolled into an end-to-end trainable deep neural network. We evaluate the proposed network on complex valued multi-coil knee images for both 6-fold and 8-fold acceleration factors, and compare with both variational and deep learning based reconstruction algorithms. The numerical results demonstrate that our method provides better reconstruction accuracy and perceptual quality by making using of the dual domain information. Clinical relevance: This improves the reconstruction quality for accelerated parallel MRI data both visually and quantitatively.

Comparison of a Deep Learning-Based Reconstruction Algorithm with Filtered Back Projection and Iterative Reconstruction Algorithms for Pediatric Abdominopelvic CT.

ObjectiveTo compare a deep learning-based reconstruction (DLR) algorithm for pediatric abdominopelvic computed tomography (CT) with filtered back projection (FBP) and iterative reconstruction (IR) algorithms.Materials and MethodsPost-contrast abdominopelvic CT scans obtained from 120 pediatric patients (mean age ± standard deviation, 8.7 ± 5.2 years; 60 males) between May 2020 and October 2020 were evaluated in this retrospective study. Images were reconstructed using FBP, a hybrid IR algorithm (ASiR-V) with blending factors of 50% and 100% (AV50 and AV100, respectively), and a DLR algorithm (TrueFidelity) with three strength levels (low, medium, and high). Noise power spectrum (NPS) and edge rise distance (ERD) were used to evaluate noise characteristics and spatial resolution, respectively. Image noise, edge definition, overall image quality, lesion detectability and conspicuity, and artifacts were qualitatively scored by two pediatric radiologists, and the scores of the two reviewers were averaged. A repeated-measures analysis of variance followed by the Bonferroni post-hoc test was used to compare NPS and ERD among the six reconstruction methods. The Friedman rank sum test followed by the Nemenyi-Wilcoxon-Wilcox all-pairs test was used to compare the results of the qualitative visual analysis among the six reconstruction methods.ResultsThe NPS noise magnitude of AV100 was significantly lower than that of the DLR, whereas the NPS peak of AV100 was significantly higher than that of the high- and medium-strength DLR (p < 0.001). The NPS average spatial frequencies were higher for DLR than for ASiR-V (p < 0.001). ERD was shorter with DLR than with ASiR-V and FBP (p < 0.001). Qualitative visual analysis revealed better overall image quality with high-strength DLR than with ASiR-V (p < 0.001).ConclusionFor pediatric abdominopelvic CT, the DLR algorithm may provide improved noise characteristics and better spatial resolution than the hybrid IR algorithm.

Deep learning reconstruction algorithm for coronary CT angiography in assessing obstructive coronary artery disease caused by calcified lesions: the clinical application value

Objective: To investigate the image quality of coronary CT angiography (CCTA) subjected to deep learning-based reconstruction algorithm (DLR) method and its diagnostic performance for stenosis caused by coronary calcified lesions. Methods: We enrolled 33 consecutive patients with known or suspected coronary artery disease (CAD) who underwent CCTA and subsequently invasive coronary angiography (ICA) within 1 month in the department of radiology, Peking Union Medical College Hospital between February 2020 and February 2021. Among them, there are 26 males and 7 females, age range from 45 to 86 (61.9±9.0) years. The CCTA images were reconstructed with DLR and hybrid iterative reconstruction (HIR). Image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) were calculated on the aorta root, left main artery, proximal left anterior descending, left circumflex, and right coronary artery of the CCTA images and were used to evaluate the objective image quality (IQ). Subjective IQ score was graded using Likert four-point scale (1 for excellent and 4 for poor). The diagnostic performance of obstructive coronary artery disease caused by calcified lesions on CCTA subjected to DLR and HIR methods were evaluated using ICA as the reference standard. Results: A total of 123 lesions in 33 patients were included in the analysis. Image noise of DLR image was significantly lower than that on HIR image(defined as the standard deviation of the attenuation values in the aortic root: 18.12±3.66 vs 24.19±5.71, P<0.001), CNR and SNR of DLR image in the aortic root were higher (CNR:43.83±23.73 vs 26.38±9.69, P<0.001,SNR:26.66±7.83 vs 21.23±8.65, P<0.001). Subjective scores of DLR was better than HIR image (1.12±0.41 vs 1.46±0.60,P<0.001). The sensitivity, specificity and accuracy of DLR and HIR images for diagnosing obstructive coronary artery disease caused by calcified lesions were 100.0%, 77.4%, 78.9% and 100.0%, 63.5%, 65.9%%, respectively. The number of false positive cases on DLR image decreased by 38% compared with HIR. Conclusions: Artificial intelligence based DLR can significantly reduce the image noise and improve the image quality of CCTA. DLR helps to improve the diagnostic performance of CCTA in assessing obstructive coronary artery disease caused by calcified lesions, which may have good clinical application value.

Deep learning based multiplexed sensitivity-encoding (DL-MUSE) for high-resolution multi-shot DWI

PurposeA phase correction method for high-resolution multi-shot (MSH) diffusion weighted imaging (DWI) is proposed. The efficacy and generalization capability of the method were validated on both healthy volunteers and patients. Theory and methodsConventionally, inter-shot phase variations for MSH echo-planar imaging (EPI) DWI are corrected by model-based algorithms. However, many acquisition imperfections are hard to measure accurately for conventional model-based methods, making the phase estimation and artifacts suppression unreliable. We propose a deep learning multiplexed sensitivity-encoding (DL-MUSE) framework to improve the phase estimations based on convolutional neural network (CNN) reconstruction. Aliasing-free single-shot (SSH) DW images, which have been used routinely in clinical settings, were used for training before the aliasing correction of MSH-DWI images. A dual-channel U-net comprising multiple convolutional layers was used for the phase estimation of MSH-EPI. The network was trained on a dataset containing 30 healthy volunteers and tested on another dataset of 52 healthy subjects and 15 patients with lesions or tumors with different shot numbers (4, 6 and 8). To further validate the generalization capability of our network, we acquired a dataset with different numbers of shots, TEs, partial Fourier factors, resolutions, ETLs, FOVs, coil numbers, and image orientations from two sites. We also compared the reconstruction performance of our proposed method with that of the conventional MUSE and SSH-EPI qualitatively and quantitatively. ResultsOur results show that DL-MUSE is capable of correcting inter-shot phase errors with high and robust performance. Compared to conventional model-based MUSE, our method, by applying deep learning-based phase corrections, showed reduced distortion, noise level, and signal loss in high b-value DWIs. The improvements of image quality become more evident as the shot number increases from 4 to 8, especially in those central regions of the images, where g-factor artifacts are severe. Furthermore, the proposed method could provide the information about the orientation of the white matter with better consistency and achieve finer fibers delineation compared to the SSH-EPI method. Besides, the experiments on volunteers and patients from two different sites demonstrated the generalizability of our proposed method preliminarily. ConclusionA deep learning-based reconstruction algorithm for MSH-EPI images, which helps improve image quality greatly, was proposed. Results from healthy volunteers and tumor patients demonstrated the feasibility and generalization performances of our method for high-resolution MSH-EPI DWI, which can be used for routine clinical applications as well as neuroimaging research.

Diagnostic performance and image quality of deep learning image reconstruction (DLIR) on unenhanced low-dose abdominal CT for urolithiasis.

Patients with urolithiasis undergo radiation overexposure from computed tomography (CT) scans. Improvement of image reconstruction is necessary for radiation dose reduction. To evaluate a deep learning-based reconstruction algorithm for CT (DLIR) in the detection of urolithiasis at low-dose non-enhanced abdominopelvic CT. A total of 75 patients who underwent low-dose abdominopelvic CT for urolithiasis were retrospectively included. Each examination included three reconstructions: DLIR; filtered back projection (FBP); and hybrid iterative reconstruction (IR; ASiR-V 70%). Image quality was subjectively and objectively assessed using attenuation and noise measurements in order to calculate the signal-to-noise ratio (SNR), absolute contrast, and contrast-to-noise ratio (CNR). Attenuation of the largest stones were also compared. Detectability of urinary stones was assessed by two observers. Image noise was significantly reduced with DLIR: 7.2 versus 17 and 22 for ASiR-V 70% and FBP, respectively. Similarly, SNR and CNR were also higher compared to the standard reconstructions. When the structures had close attenuation values, contrast was lower with DLIR compared to ASiR-V. Attenuation of stones was also lowered in the DLIR series. Subjective image quality was significantly higher with DLIR. The detectability of all stones and stones >3 mm was excellent with DLIR for the two observers (intraclass correlation [ICC] = 0.93 vs. 0.96 and 0.95 vs. 0.99). For smaller stones (<3 mm), results were different (ICC = 0.77 vs. 0.86). For low-dose abdominopelvic CT, DLIR reconstruction exhibited image quality superior to ASiR-V and FBP as well as an excellent detection of urinary stones.

Deep learning reconstruction of digital breast tomosynthesis images for accurate breast density and patient-specific radiation dose estimation.

The two-dimensional nature of mammography makes estimation of the overall breast density challenging, and estimation of the true patient-specific radiation dose impossible. Digital breast tomosynthesis (DBT), a pseudo-3D technique, is now commonly used in breast cancer screening and diagnostics. Still, the severely limited 3rd dimension information in DBT has not been used, until now, to estimate the true breast density or the patient-specific dose. This study proposes a reconstruction algorithm for DBT based on deep learning specifically optimized for these tasks. The algorithm, which we name DBToR, is based on unrolling a proximal-dual optimization method. The proximal operators are replaced with convolutional neural networks and prior knowledge is included in the model. This extends previous work on a deep learning-based reconstruction model by providing both the primal and the dual blocks with breast thickness information, which is available in DBT. Training and testing of the model were performed using virtual patient phantoms from two different sources. Reconstruction performance, and accuracy in estimation of breast density and radiation dose, were estimated, showing high accuracy (density <±3%; dose <±20%) without bias, significantly improving on the current state-of-the-art. This work also lays the groundwork for developing a deep learning-based reconstruction algorithm for the task of image interpretation by radiologists.

Noise Reduction Effect of Deep-learning-based Image Reconstruction Algorithms in Thin-section Chest CT

To evaluate the noise reduction effect of deep learning-based reconstruction algorithms in thin-section chest CT images by analyzing images reconstructed with filtered back projection (FBP), adaptive statistical iterative reconstruction (ASIR), and deep learning image reconstruction (DLIR) algorithms. The chest CT scan raw data of 47 patients were included in this study. Images of 0.625 mm were reconstructed using six reconstruction methods, including FBP, ASIR hybrid reconstruction (ASIR50%, ASIR70%), and deep learning low, medium and high modes (DL-L, DL-M, and DL-H). After the regions of interest were outlined in the aorta, skeletal muscle and lung tissue of each group of images, the CT values, SD values and signal-to-noise ratio (SNR) of the regions of interest were measured, and two radiologists evaluated the image quality. CT values, SD values and SNR of the images obtained by the six reconstruction methods showed statistically significant difference ( P<0.001). There were statistically significant differences in the image quality scores of the six reconstruction methods ( P<0.001). Images reconstruced with DL-H have the lowest noise and the highest overall quality score. The model based on deep learning can effectively reduce the noise of thin-section chest CT images and improve the image quality. Among the three deep-learning models, DL-H showed the best noise reduction effect.

Deep learning-based reconstruction in ultra-high-resolution computed tomography: Can image noise caused by high definition detector and the miniaturization of matrix element size be improved?

This study aimed to assess the noise characteristics of ultra-high-resolution computed tomography (UHRCT) with deep learning-based reconstruction (DLR). Two different diameters of water phantom were scanned with three different resolution acquisition modes. Images were reconstructed by filtered back projection (FBP), hybrid iterative reconstruction (hybrid-IR), and DLR. Image noise analysis was performed with noise magnitude, peak frequency (fp) of the noise power spectrum (NPS), and the square root of the area under the curve (√AUCNPS) for the NPS curve. The noise magnitude was up to 3.30 times higher for the FBP acquired in SHR mode than that for the NR mode. The fp values of the FBP were 0.20-0.21, 0.34-0.36, and 0.34-0.37 cycles/mm for normal resolution (NR), high resolution (HR), and super high resolution (SHR) mode, respectively. The fp of hybrid-IR was 0.16-0.19, 0.21-0.26, and 0.23-0.26 cycles/mm for NR, HR, and SHR mode, respectively. The fp of DLR was 0.21-0.32 and 0.22-0.33 cycles/mm for HR and SHR mode, respectively. √AUCNPS showed that the highest value in FBP images of the SHR mode was up to 1.89 times that of the NR mode. DLR in the HR and SHR modes showed high noise reduction while suppressing fp shift with respect to FBP. The new DLR algorithm could be a solution to the noise increase due to the high-definition detector elements and the small reconstruction matrix element size.

Technical Improvements in Head and Neck MR Imaging: At the Cutting Edge.

Head and neck MR imaging is technically challenging because of magnetic field inhomogeneity, respiratory and swallowing motion, and necessity of high-resolution imaging to trace key anatomic structures. These challenges have been answered by advances in MR imaging technology, including isovolumetric three-dimensional imaging, robust fat-water separation techniques, and novel deep learning-based reconstruction algorithms. New applications of MR imaging have been advanced and functional imaging has been improved. Improvements in acquisition and reconstruction technique facilitate novel applications of morphologic and functional imaging. This results in opportunities to improve diagnosis, staging, and treatment selection through application of advanced MR imaging techniques.

DR2-Net: Deep Residual Reconstruction Network for image compressive sensing

Most traditional algorithms for compressive sensing image reconstruction suffer from the intensive computation. Recently, deep learning-based reconstruction algorithms have been reported, which dramatically reduce the time complexity than iterative reconstruction algorithms. In this paper, we propose a novel Deep Residual Reconstruction Network (DR2-Net) to reconstruct the image from its Compressively Sensed (CS) measurement. The DR2-Net is proposed based on two observations: (1) linear mapping could reconstruct a high-quality preliminary image, and (2) residual learning could further improve the reconstruction quality. Accordingly, DR2-Net consists of two components, i.e., linear mapping network and residual network, respectively. Specifically, the fully-connected layer in neural network implements the linear mapping network. We then expand the linear mapping network to DR2-Net by adding several residual learning blocks to enhance the preliminary image. Extensive experiments demonstrate that the DR2-Net outperforms traditional iterative methods and recent deep learning-based methods by large margins at measurement rates 0.01, 0.04, 0.1, and 0.25, respectively. The code of DR2-Net has been released on: https://github.com/coldrainyht/caffe_dr2.