Deep Learning-based Image Reconstruction Research Articles

The impact of deep learning reconstruction on image quality and coronary CT angiography-derived fractional flow reserve values.

To explore the impact of deep learning reconstruction (DLR) on image quality and machine learning-based coronary CT angiography (CTA)-derived fractional flow reserve (CT-FFRML) values. Thirty-three consecutive patients with known or suspected coronary artery disease who underwent coronary CTA and subsequent invasive coronary angiography were enrolled. DLR was compared with filtered back projection (FBP), statistical-based iterative reconstruction (SBIR), model-based iterative reconstruction (MBIR) Cardiac, and MBIR Cardiac sharp for objective image qualities of coronary CTA. Invasive fractional flow reserve (FFR) and quantitative flow ratio (QFR) were used as the reference standards. The diagnostic performances of different reconstruction approach-based CT-FFRML were calculated. A total of 182 lesions in 33 patients were enrolled for analysis. The image quality of DLR was superior to the others. There were no significant differences in the CT-FFRML values among these five approaches (all p > 0.05). Of the 182 lesions, 17 had invasive FFR results, and 70 had QFR results. Using FFR as a reference, MBIR Cardiac, MBIR Cardiac sharp, and DLR achieved equal diagnostic performance, slightly higher than the other reconstruction approaches (MBIR Cardiac, MBIR Cardiac sharp, and DLR: AUC = 0.82, FBP and AIDR: AUC = 0.78, all p > 0.05). Using QFR as a reference, the AUCs of FBP, SBIR, MBIR Cardiac, MBIR Cardiac sharp, and DLR were 0.83, 0.81, 0.86, 0.84, and 0.83, respectively (all p > 0.05). Our study showed that the DLR algorithm improved image quality, but there were no significant differences in the CT-FFRML values and diagnostic performance among different reconstruction approaches. • Deep learning-based image reconstruction (DLR) improves the image quality of coronary CTA. • CT-FFRML values and diagnostic performance of DLR revealed no significant differences compared to other reconstruction approaches.

Adversarial Resolution Enhancement for Electrical Capacitance Tomography Image Reconstruction.

High-quality image reconstruction is essential for many electrical capacitance tomography (CT) applications. Raw capacitance measurements are used in the literature to generate low-resolution images. However, such low-resolution images are not sufficient for proper functionality of most systems. In this paper, we propose a novel adversarial resolution enhancement (ARE-ECT) model to reconstruct high-resolution images of inner distributions based on low-quality initial images, which are generated from the capacitance measurements. The proposed model uses a UNet as the generator of a conditional generative adversarial network (CGAN). The generator’s input is set to the low-resolution image rather than the typical random input signal. Additionally, the CGAN is conditioned by the input low-resolution image itself. For evaluation purposes, a massive ECT dataset of 320 K synthetic image–measurement pairs was created. This dataset is used for training, validating, and testing the proposed model. New flow patterns, which are not exposed to the model during the training phase, are used to evaluate the feasibility and generalization ability of the ARE-ECT model. The superiority of ARE-ECT, in the efficient generation of more accurate ECT images than traditional and other deep learning-based image reconstruction algorithms, is proved by the evaluation results. The ARE-ECT model achieved an average image correlation coefficient of more than and an average relative image error about .

Impact of deep learning-based image reconstruction on image quality compared with adaptive statistical iterative reconstruction-Veo in renal and adrenal computed tomography.

OBJECTIVE: To evaluate image quality of deep learning-based image reconstruction (DLIR) in contrast-enhanced renal and adrenal computed tomography (CT) compared with adaptive statistical iterative reconstruction-Veo (ASiR-V).METHODS: We prospectively recruited 52 patients. All images were reconstructed with ASiR-V 30%, ASiR-V 70%, and DLIR at low, medium, and high reconstruction strengths. CT number, noise, noise reduction rate, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) were measured and calculated within the region of interest (ROI) on subcutaneous fat, bilateral renal cortices, renal medulla, renal arteries, and adrenal glands. For qualitative analyses, the differentiation of the renal cortex and medulla, conspicuity of the adrenal gland boundary, sharpness, artifacts, and subjective noise were assessed. The overall image quality was calculated on a scale from 0 (worst) to 15 (best) based on the five values above and the score≥9 was acceptable.RESULTS: CT number does not significantly differ between the reconstruction datasets. Noise does not significantly differ between ASiR-V 30% and DLIR-L, but it is significantly lower using ASiR-V 70%, DLIR-M, and DLIR-H. The noise reduction rate relative to ASiR-V 30% is significantly different between the DLIR groups and ASiR-V 70%, and DLIR-H yields the highest noise reduction rate (61.6%). SNR and CNR are higher for DLIR-M, DLIR-H, and ASiR-V 70% than for ASiR-V 30% and DLIR-L. DLIR-H shows the best SNR and CNR. The overall image quality yields the same pattern for DLIR-H, with the highest score. Percentages of cases with overall image quality score≥9 are 100% (DLIR-H), 94.23% (DLIR-M), 90.38% (ASiR-V70%), 67.31% (DLIR-L), and 63.46% (ASiR-V30%), respectively.CONCLUSIONS: DLIR significantly improved the objective and subjective image quality of renal and adrenal CTs, yielding superior noise reduction compared with ASiR-V.

Evaluation of Abdominal CT Obtained Using a Deep Learning-Based Image Reconstruction Engine Compared with CT Using Adaptive Statistical Iterative Reconstruction.

Purpose:To compare the image quality of CT obtained using a deep learning-based image reconstruction (DLIR) engine with images with adaptive statistical iterative reconstruction-V (AV).Materials and Methods:Using a phantom, the noise power spectrum (NPS) and task-based transfer function (TTF) were measured in images with different reconstructions (filtered back projection [FBP], AV30, 50, 100, DLIR-L, M, H) at multiple doses. One hundred and twenty abdominal CTs with 30% dose reduction were processed using AV30, AV50, DLIR-L, M, H. Objective and subjective analyses were performed.Results:The NPS peak of DLIR was lower than that of AV30 or AV50. Compared with AV30, the NPS average spatial frequencies were higher with DLIR-L or DLIR-M. For lower contrast objects, TTF in images with DLIR were higher than those with AV. The standard deviation in DLIR-H and DLIR-M was significantly lower than AV30 and AV50. The overall image quality was the best for DLIR-M (p < 0.001).Conclusions:DLIR showed improved image quality and decreased noise under a decreased radiation dose.

Unenhanced abdominal low-dose CT reconstructed with deep learning-based image reconstruction: image quality and anatomical structure depiction

PurposeTo evaluate the utility of deep learning-based image reconstruction (DLIR) algorithm in unenhanced abdominal low-dose CT (LDCT).Materials and methodsTwo patient groups were included in this prospective study: 58 consecutive patients who underwent unenhanced abdominal standard-dose CT reconstructed with hybrid iterative reconstruction (SDCT group) and 48 consecutive patients who underwent unenhanced abdominal LDCT reconstructed with high strength level of DLIR (LDCT group). The background noise and signal-to-noise ratio (SNR) of the liver, pancreas, spleen, kidney, abdominal aorta, inferior vena cava, and portal vein were calculated. Two radiologists qualitatively assessed the overall image noise, overall image quality, and abdominal anatomical structures depiction. Quantitative and qualitative parameters and size-specific dose estimates (SSDE) were compared between SDCT and LDCT groups.ResultsThe background noise was lower in LDCT group than in SDCT group (P = 0.02). SNRs were higher in LDCT group than in SDCT group (P < 0.001–0.004) except for the liver. Overall image noise was superior in LDCT group than in SDCT group (P < 0.001). Overall image quality was not different between SDCT and LDCT groups (P = 0.25–0.26). Depiction of almost all abdominal anatomical structures was equal to or better in LDCT group than in SDCT group (P < 0.001–0.88). The SSDE was lower in LDCT group (4.0 mGy) than in SDCT group (20.6 mGy) (P < 0.001).ConclusionsDLIR facilitates substantial radiation dose reduction of > 75% and significantly reduces background noise. DLIR can maintain image quality and anatomical structure depiction in unenhanced abdominal LDCT.

Task-specific spatial resolution properties of iterative and deep learning-based reconstructions in computed tomography: Comparison using tasks assuming small and large enhanced vessels

PurposeThe present study aims to evaluate TTFs of deep-learning-based image reconstruction (DLIR) and iterative reconstruction (IR) in computed tomography (CT) using a conventional task with a rod object with a diameter of 30 mm and a newly-proposed task with a wire of 1 mm in diameter, simulating large and small enhanced vessels, respectively. MethodsThe rod or wire phantom made of a material equivalent to diluted iodine that exhibits about 270 Hounsfield unit (HU) was placed inside a 30-cm water phantom. In-plane and z-directional TTFs were measured for the rod using the circular edge (CE) and plane edge (PE) methods, respectively. By using the wire (iodine wire: IW), in-plane and z-directional TTFs were measured using Fourier transform (IW method). TTFs of filtered back projection (FBP), IR, and DLIR of a 256-row CT system and FBP and IR of a 64-row CT system were evaluated with CT dose indices of 10 and 5 mGy. ResultsFor DLIR and IR, TTFs measured using the IW method were notably lower than those using the CE (or PE) method; moreover, they were also lower than those of corresponding FBP, indicating that the small enhanced vessels with a diameter of about 1 mm would be blurred with both DLIR and IR. ConclusionsThe proposed IW method has turned out to be effective to evaluate TTFs for small enhanced vessels, which have not been properly evaluated by the CE or PE method conventionally recommended.

Uncertainty Quantification for Deep Unrolling-Based Computational Imaging

Deep unrolling is an emerging deep learning-based image reconstruction methodology that bridges the gap between model-based and purely deep learning-based image reconstruction methods. Although deep unrolling methods achieve state-of-the-art performance for imaging problems and allow the incorporation of the observation model into the reconstruction process, they do not provide any uncertainty information about the reconstructed image, which severely limits their use in practice, especially for safety-critical imaging applications. In this paper, we propose a learning-based image reconstruction framework that incorporates the observation model into the reconstruction task and that is capable of quantifying epistemic and aleatoric uncertainties, based on deep unrolling and Bayesian neural networks. We demonstrate the uncertainty characterization capability of the proposed framework on magnetic resonance imaging and computed tomography reconstruction problems. We investigate the characteristics of the epistemic and aleatoric uncertainty information provided by the proposed framework to motivate future research on utilizing uncertainty information to develop more accurate, robust, trustworthy, uncertainty-aware, learning-based image reconstruction and analysis methods for imaging problems. We show that the proposed framework can provide uncertainty information while achieving comparable reconstruction performance to state-of-the-art deep unrolling methods.

Evaluation of Ultra-low-dose (ULD) Lung Computed Tomography (CT) Using Deep-learning: A Phantom Study

As an electromagnetic wave, X-rays are used to acquire diagnostic CT images. The aim of this phantom study was to evaluate the image quality of ultra-low-dose (ULD) lung computed tomography (CT) achieved using a deep-learning based image reconstruction method. The chest phantom was scanned with a tube voltage of 100 kV for various CT dose index (CTDIvol) conditions: 0.4 mGy for ultra-low-dose (ULD), 0.6 mGy for low-dose (LD), 2.7 mGy for standard (SD), and 7.1 mGy for large size (LS). The signal-to-noise ratio (SNR) and noise values in reconstructions produced via filtered back projection (FBP), iterative reconstruction (IR), and deep convolutional neural network (DCNN) were computed for comparison. The quantitative results of both the SNR and noise indicate that the adoption of the DCNN makes the image reconstruction in the ULD setting more stable and robust, achieving a higher image quality when compared with the FBP algorithm in the SD condition. Compared with the conventional FBP and IR, the proposed deep learning-based image reconstruction approach can improve the ULD CT image quality while significantly reducing the patient dose.

An Open-Source Deep Learning Network for Reconstruction of High-Resolution OCT Angiograms of Retinal Intermediate and Deep Capillary Plexuses.

PurposeWe propose a deep learning–based image reconstruction algorithm to produce high-resolution optical coherence tomographic angiograms (OCTA) of the intermediate capillary plexus (ICP) and deep capillary plexus (DCP).MethodsIn this study, 6-mm × 6-mm macular scans with a 400 × 400 A-line sampling density and 3-mm × 3-mm scans with a 304 × 304 A-line sampling density were acquired on one or both eyes of 180 participants (including 230 eyes with diabetic retinopathy and 44 healthy controls) using a 70-kHz commercial OCT system (RTVue-XR; Optovue, Inc., Fremont, California, USA). Projection-resolved OCTA algorithm removed projection artifacts in voxel. ICP and DCP angiograms were generated by maximum projection of the OCTA signal within the respective plexus. We proposed a deep learning–based method, which receives inputs from registered 3-mm × 3-mm ICP and DCP angiograms with proper sampling density as the ground truth reference to reconstruct 6-mm × 6-mm high-resolution ICP and DCP en face OCTA. We applied the same network on 3-mm × 3-mm angiograms to enhance these images further. We evaluated the reconstructed 3-mm × 3-mm and 6-mm × 6-mm angiograms based on vascular connectivity, Weber contrast, false flow signal (flow signal erroneously generated from background), and the noise intensity in the foveal avascular zone.ResultsCompared to the originals, the Deep Capillary Angiogram Reconstruction Network (DCARnet)–enhanced 6-mm × 6-mm angiograms had significantly reduced noise intensity (ICP, 7.38 ± 25.22, P < 0.001; DCP, 11.20 ± 22.52, P < 0.001), improved vascular connectivity (ICP, 0.95 ± 0.01, P < 0.001; DCP, 0.96 ± 0.01, P < 0.001), and enhanced Weber contrast (ICP, 4.25 ± 0.10, P < 0.001; DCP, 3.84 ± 0.84, P < 0.001), without generating false flow signal when noise intensity lower than 650. The DCARnet-enhanced 3-mm × 3-mm angiograms also reduced noise, improved connectivity, and enhanced Weber contrast in 3-mm × 3-mm ICP and DCP angiograms from 101 eyes. In addition, DCARnet preserved the appearance of the dilated vessels in the reconstructed angiograms in diabetic eyes.ConclusionsDCARnet can enhance 3-mm × 3-mm and 6-mm × 6-mm ICP and DCP angiogram image quality without introducing artifacts.Translational RelevanceThe enhanced 6-mm × 6-mm angiograms may be easier for clinicians to interpret qualitatively.

Deep Learning-Based Image Reconstruction for CT Angiography of the Aorta.

To evaluate the impact of a novel, deep-learning-based image reconstruction (DLIR) algorithm on image quality in CT angiography of the aorta, we retrospectively analyzed 51 consecutive patients who underwent ECG-gated chest CT angiography and non-gated acquisition for the abdomen on a 256-dectector-row CT. Images were reconstructed with adaptive statistical iterative reconstruction (ASIR-V) and DLIR. Intravascular image noise, the signal-to-noise ratio (SNR) and the contrast-to-noise ratio (CNR) were quantified for the ascending aorta, the descending thoracic aorta, the abdominal aorta and the iliac arteries. Two readers scored subjective image quality on a five-point scale. Compared to ASIR-V, DLIR reduced the median image noise by 51–54% for the ascending aorta and the descending thoracic aorta. Correspondingly, median CNR roughly doubled for the ascending aorta and descending thoracic aorta. There was a 38% reduction in image noise for the abdominal aorta and the iliac arteries, with a corresponding improvement in CNR. Median subjective image quality improved from good to excellent at all anatomical levels. In CT angiography of the aorta, DLIR substantially improved objective and subjective image quality beyond what can be achieved by state-of-the-art iterative reconstruction. This can pave the way for further radiation or contrast dose reductions.

Diagnostic Improvements of Deep Learning-Based Image Reconstruction for Assessing Calcification-Related Obstructive Coronary Artery Disease.

Objectives: The objective of this study was to explore the diagnostic value of deep learning-based image reconstruction (DLR) and hybrid iterative reconstruction (HIR) for calcification-related obstructive coronary artery disease (CAD) evaluation by using coronary CT angiography (CCTA) images and subtraction CCTA images.Methods: Forty-two consecutive patients with known or suspected coronary artery disease who underwent coronary CTA on a 320-row CT scanner and subsequent invasive coronary angiography (ICA), which was used as the reference standard, were enrolled. The DLR and HIR images were reconstructed as CTADLR and CTAHIR, and, based on which, the corresponding subtraction CCTA images were established as CTAsDLR and CTAsHIR, respectively. Qualitative images quality comparison was performed by using a Likert 4 stage score, and quantitative images quality parameters, including image noise, signal-to-noise ratio, and contrast-to-noise ratio were calculated. Diagnostic performance on the lesion level was assessed and compared among the four CCTA approaches (CTADLR, CTAHIR, CTAsDLR, and CTAsHIR).Results: There were 166 lesions of 86 vessels in 42 patients (32 men and 10 women; 62.9 ± 9.3 years) finally enrolled for analysis. The qualitative and quantitative image qualities of CTAsDLR and CTADLR were superior to those of CTAsHIR and CTAHIR, respectively. The diagnostic accuracies of CTAsDLR, CTADLR, CTAsHIR, and CTAHIR to identify calcification-related obstructive diameter stenosis were 83.73%, 69.28%, 75.30%, and 65.66%, respectively. The false-positive rates of CTAsDLR, CTADLR, CTAsHIR, and CTAHIR for luminal diameter stenosis ≥50% were 15%, 31%, 24%, and 34%, respectively. The sensitivity and the specificity to identify ≥50% luminal diameter stenosis was 90.91% and 83.23% for CTAsDLR.Conclusion: Our study showed that deep learning–based image reconstruction could improve the image quality of CCTA images and diagnostic performance for calcification-related obstructive CAD, especially when combined with subtraction technique.

Deep learning based image reconstruction for sparse-view diffuse optical tomography

Three critical problems, including high-cost instrumentations, time-consuming image recovery, and low image quality, limit clinical applications of diffuse optical tomography (DOT). Image reconstruction based on deep learning can enhance the image quality of DOT, especially in the case of low number of measurements where image reconstruction becomes a more ill-posed and underdetermined problem. Here, we present a sparse-view image reconstruction based on deep learning to recover the absorption coefficient of a phantom. The presented neural network enhances image quality and the results show that deep learning can enhance the contrast to noise-ratio effectively (more than 80%). The presented method provides recovered images with more accurate localization capabilities (an increase in localization metric by more than 30% compared with the classic method). Moreover, the time of image reconstruction is reduced by three orders of magnitude. This method can be applied for single-source DOT that significantly reduces the cost of required instruments for brain imaging and cervical or thyroid cancer screening.

Dose reduction potential of vendor-agnostic deep learning model in comparison with deep learning-based image reconstruction algorithm on CT: a phantom study.

ObjectivesTo compare the dose reduction potential (DRP) of a vendor-agnostic deep learning model (DLM, ClariCT.AI) with that of a vendor-specific deep learning–based image reconstruction algorithm (DLR, TrueFidelity™).MethodsComputed tomography (CT) images of a multi-sized image quality phantom (Mercury v4.0) were acquired under six radiation dose levels (0.48/0.97/1.93/3.87/7.74/15.47 mGy) and were reconstructed using filtered back projection (FBP) and three strength levels of the DLR (low/medium/high). The FBP images were denoised using the DLM. For all DLM and DLR images, the detectability index (d′) (a task-based detection performance metric) was obtained, under various combinations of three target sizes (10/5/1 mm), five inlets (CT value difference with the background; −895/50/90/335/1000 HU), five phantom diameters (36/31/26/21/16 cm), and six radiation dose levels. Dose reduction potential (DRP) measures the dose reduction made by using DLM or DLR, while yielding d′ equivalent to that of FBP at full dose.ResultsThe DRPs of the DLM, DLR-low, DLR-medium, and DLR-high were 86% (81–88%), 60% (46–67%), 76% (60–81%), and 87% (78–92%), respectively. For 10-mm targets, the DRP of the DLM (87%) was higher than that of all DLR algorithms (58–86%). However, for smaller targets (5 mm/1 mm), the DRPs of the DLR-high (89/88%) were greater than those of the DLM (87/84%).ConclusionThe dose reduction potential of the vendor-agnostic DLM was shown to be comparable to that of the vendor-specific DLR at high strength and superior to those of the DLRs at medium and low strengths.Key Points• DRP of the vendor-agnostic model was comparable to that of high-strength vendor-specific model and superior to those of medium- and low-strength models.• Under various radiation dose levels, the deep learning model shows higher detectability indexes compared to FBP.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00330-021-08199-9.

Deep learning-based image reconstruction for TOF PET with DIRECT data partitioning format

Conventional positron emission tomography (PET) image reconstruction is achieved by the statistical iterative method. Deep learning provides another opportunity for speeding up the image reconstruction process. However, conventional deep learning-based image reconstruction requires a fully connected network for learning the Radon transform. The use of fully connected networks greatly complicated the network and increased hardware cost. In this study, we proposed a novel deep learning-based image reconstruction method by utilizing the DIRECT data partitioning method. The U-net structure with only convolutional layers was used in our approach. Patch-based model training and testing were used to achieve 3D reconstructions within current hardware limitations. Time-of-flight (TOF)-histoimages were first generated from the listmode data to replace conventional sinograms. Different projection angles were used as different channels in the input. A total of 15 patient data were used in this study. For each patient, the dynamic whole-body scanning protocol was used to expand the training dataset and a total of 372 separate scans were included. The leave-one-patient-out validation method was used. Two separate studies were carried out. In the first study, the measured TOF-histoimages were directly used for model training and testing, to study the performance of the method in real-world applications. In the second study, TOF-histoimages were simulated from already reconstructed images to exclude the scatters, randoms, attenuation-activity mismatch effects. This study was used to evaluate the optimal performance when all other corrections are ideal. Volumes of interests were placed in the liver and lesion region to study image noise and lesion quantitations. The reconstructed images using the proposed deep learning method showed similar image quality when compared with the conventional expectation-maximization approach. A minimal difference was observed when the simulated TOF-histoimages were used as model input and testing, suggesting the deep learning model can indeed learn the reconstruction process. Some quantitative difference was observed when the measured TOF-histoimages were used. The two studies suggested that the major difference is caused by inaccurate corrections performed by the network itself, which indicated that physics-based corrections are still required for better quantitative performance. In conclusion, we have proposed a novel deep learning-based image reconstruction method for TOF PET. With the help of the DIRECT data partitioning method, no fully connected layers were used and 3D image reconstruction can be directly achieved within the limits of the current hardware.

Neural network-based image reconstruction in swept-source optical coherence tomography using undersampled spectral data

Optical coherence tomography (OCT) is a widely used non-invasive biomedical imaging modality that can rapidly provide volumetric images of samples. Here, we present a deep learning-based image reconstruction framework that can generate swept-source OCT (SS-OCT) images using undersampled spectral data, without any spatial aliasing artifacts. This neural network-based image reconstruction does not require any hardware changes to the optical setup and can be easily integrated with existing swept-source or spectral-domain OCT systems to reduce the amount of raw spectral data to be acquired. To show the efficacy of this framework, we trained and blindly tested a deep neural network using mouse embryo samples imaged by an SS-OCT system. Using 2-fold undersampled spectral data (i.e., 640 spectral points per A-line), the trained neural network can blindly reconstruct 512 A-lines in 0.59 ms using multiple graphics-processing units (GPUs), removing spatial aliasing artifacts due to spectral undersampling, also presenting a very good match to the images of the same samples, reconstructed using the full spectral OCT data (i.e., 1280 spectral points per A-line). We also successfully demonstrate that this framework can be further extended to process 3× undersampled spectral data per A-line, with some performance degradation in the reconstructed image quality compared to 2× spectral undersampling. Furthermore, an A-line-optimized undersampling method is presented by jointly optimizing the spectral sampling locations and the corresponding image reconstruction network, which improved the overall imaging performance using less spectral data points per A-line compared to 2× or 3× spectral undersampling results. This deep learning-enabled image reconstruction approach can be broadly used in various forms of spectral-domain OCT systems, helping to increase their imaging speed without sacrificing image resolution and signal-to-noise ratio.

Parametric comparison between sparsity-based and deep learning-based image reconstruction of super-resolution fluorescence microscopy.

Sparsity-based and deep learning-based image reconstruction algorithms are two promising approaches to accelerate the image acquisition process for localization-based super-resolution microscopy, by allowing a higher density of fluorescing emitters to be imaged in a single frame. Despite the surging popularity, a comprehensive parametric study guiding the practical applications of sparsity-based and deep learning-based image reconstruction algorithms is yet to be conducted. In this study, we examined the performance of sparsity- and deep learning-based algorithms in reconstructing super-resolution images using simulated fluorescent microscopy images. The simulated images were synthesized with varying levels of sparsity and connectivity. We found the deep learning-based VDSR recovers image faster, with a higher recall rate and localization accuracy. The sparsity-based SPIDER recovers more zero pixels truthfully. We also compared the two algorithms using images acquired from a real super-resolution experiment, yielding results agreeing with the results from the evaluation using simulated images. We concluded that VDSR is preferable when accurate emitter localization is needed while SPIDER is more suitable when evaluation of the number of emitters is critical.

Deep learning in magnetic resonance image reconstruction.

Magnetic resonance (MR) imaging visualises soft tissue contrast in exquisite detail without harmful ionising radiation. In this work, we provide a state-of-the-art review on the use of deep learning in MR image reconstruction from different image acquisition types involving compressed sensing techniques, parallel image acquisition and multi-contrast imaging. Publications with deep learning-based image reconstruction for MR imaging were identified from the literature (PubMed and Google Scholar), and a comprehensive description of each of the works was provided. A detailed comparison that highlights the differences, the data used and the performance of each of these works were also made. A discussion of the potential use cases for each of these methods is provided. The sparse image reconstruction methods were found to be most popular in using deep learning for improved performance, accelerating acquisitions by around 4-8 times. Multi-contrast image reconstruction methods rely on at least one pre-acquired image, but can achieve 16-fold, and even up to 32- to 50-fold acceleration depending on the set-up. Parallel imaging provides frameworks to be integrated in many of these methods for additional speed-up potential. The successful use of compressed sensing techniques and multi-contrast imaging with deep learning and parallel acquisition methods could yield significant MR acquisition speed-ups within clinical routines in the near future.

Improving the image quality of pediatric chest CT angiography with low radiation dose and contrast volume using deep learning image reconstruction.

Chest CT angiography (CTA) is a common clinical examination technique for children. Iterative reconstruction algorithms are often used to reduce image noise but encounter limitations under low dose conditions. Deep learning-based image reconstruction algorithms have been developed to overcome these limitations. We assessed the quantitative and qualitative image quality of thin-slice chest CTA in children acquired with low radiation dose and contrast volume by using a deep learning image reconstruction (DLIR) algorithm. A total of 33 children underwent chest CTA with 70 kVp and automatic tube current modulation for noise indices of 11-15 based on their age and contrast volume of 0.8-1.2 mL/kg. Images were reconstructed with 50% and 100% adaptive statistical iterative reconstruction-V (ASIR-V) and high-setting DLIR (DLIR-H) at 0.625 mm slice thickness. Two radiologists evaluated images in consensus for overall image noise, artery margin, and artery contrast separately on a 5-point scale (5, excellent; 4, good; 3, acceptable; 2, sub-acceptable, and 1, not acceptable). The CT value and image noise of the descending aorta and back muscle were measured. Radiation dose and contrast volume was recorded. The volume CT dose index, dose length product, and contrast volume were 1.37±0.29 mGy, 35.43±10.59 mGy·cm, and 25.43±13.32 mL, respectively. The image noises (in HU) of the aorta with DLIR-H (19.24±5.77) and 100% ASIR-V (20.45±6.93) were not significantly different (P>0.05) and were substantially lower than 50% ASIR-V (29.45±7.59) (P<0.001). The 100% ASIR-V images had over-smoothed artery margins, but only the DLIR-H images provided acceptable scores on all 3 aspects of the qualitative image quality evaluation. It is feasible to improve the image quality of a low radiation dose and contrast volume chest CTA in children using the high-setting DLIR algorithm.

Deep learning-based image reconstruction for few-view computed tomography

Few-view computed tomography (CT) is able to provide the internal information of an object for diagnosis, while reducing radiation dose and decreasing side effects. In spite of these benefits, few-view CT images with conventional reconstruction algorithms are suffer from structural distortion and artifacts due to insufficient projections. In this study, a deep learning-based network was proposed for reconstructing few-view CT images and improving image quality. The proposed network consisted of fully connected, convolution and deconvolution layers, and the network training was implemented to directly transform few-view projections into reconstructed images without expert reconstruction knowledge. The quality of the output images obtained from the trained network was compared with those reconstructed by using the filtered back-projection (FBP) and simultaneous algebraic reconstruction technique (SART) algorithms. The results showed that the proposed network is able to improve the quantitative accuracy and noise property of reconstructed images, and the proposed network is more suitable for few-view CT reconstruction than the other algorithms. In conclusion, the proposed network can be potentially used to improve few-view CT image quality and considered as an alternative for few-view CT reconstruction. The source code is available on a GitHub repository at https://github.com/ YIMDOBIN/CT_reconstruction_via_DL.

Adaptive convolutional neural networks for accelerating magnetic resonance imaging via k-space data interpolation.

Deep learning in k-space has demonstrated great potential for image reconstruction from undersampled k-space data in fast magnetic resonance imaging (MRI). However, existing deep learning-based image reconstruction methods typically apply weight-sharing convolutional neural networks (CNNs) to k-space data without taking into consideration the k-space data's spatial frequency properties, leading to ineffective learning of the image reconstruction models. Moreover, complementary information of spatially adjacent slices is often ignored in existing deep learning methods. To overcome such limitations, we have developed a deep learning algorithm, referred to as adaptive convolutional neural networks for k-space data interpolation (ACNN-k-Space), which adopts a residual Encoder-Decoder network architecture to interpolate the undersampled k-space data by integrating spatially contiguous slices as multi-channel input, along with k-space data from multiple coils if available. The network is enhanced by self-attention layers to adaptively focus on k-space data at different spatial frequencies and channels. We have evaluated our method on two public datasets and compared it with state-of-the-art existing methods. Ablation studies and experimental results demonstrate that our method effectively reconstructs images from undersampled k-space data and achieves significantly better image reconstruction performance than current state-of-the-art techniques. Source code of the method is available at https://gitlab.com/qgpmztmf/acnn-k-space.