Attention-based deep neural network for partial volume correction in brain 18F-FDG PET imaging

PURPOSE

This work set out to propose an attention-based deep neural network to predict partial volume corrected images from PET data not utilizing anatomical information.

METHODS

An attention-based convolutional neural network (ATB-Net) is developed to predict PVE-corrected images in brain PET imaging by concentrating on anatomical areas of the brain. The performance of the deep neural network for performing PVC without using anatomical images was evaluated for two PVC methods, including iterative Yang (IY) and reblurred Van-Cittert (RVC) approaches. The RVC and IY PVC approaches were applied to PET images to generate the reference images. The training of the U-Net network for the partial volume correction was trained twice, once without using the attention module and once with the attention module concentrating on the anatomical brain regions.

RESULTS

Regarding the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean square error (RMSE) metrics, the proposed ATB-Net outperformed the standard U-Net model (without attention compartment). For the RVC technique, the ATB-Net performed just marginally better than the U-Net; however, for the IY method, which is a region-wise method, the attention-based approach resulted in a substantial improvement. The mean absolute relative SUV difference and mean absolute relative bias improved by 38.02 % and 91.60 % for the RVC method and 77.47 % and 79.68 % for the IY method when using the ATB-Net model, respectively.

CONCLUSIONS

Our results propose that without using anatomical data, the attention-based DL model could perform PVC on PET images, which could be employed for PVC in PET imaging.

Impact of simulated reduced injected dose on the assessment of amyloid PET scans

PURPOSE

To investigate the impact of reduced injected doses on the quantitative and qualitative assessment of the amyloid PET tracers [18F]flutemetamol and [18F]florbetaben.

METHODS

Cognitively impaired and unimpaired individuals (N = 250, 36% Aβ-positive) were included and injected with [18F]flutemetamol (N = 175) or [18F]florbetaben (N = 75). PET scans were acquired in list-mode (90-110 min post-injection) and reduced-dose images were simulated to generate images of 75, 50, 25, 12.5 and 5% of the original injected dose. Images were reconstructed using vendor-provided reconstruction tools and visually assessed for Aβ-pathology. SUVRs were calculated for a global cortical and three smaller regions using a cerebellar cortex reference tissue, and Centiloid was computed. Absolute and percentage differences in SUVR and CL were calculated between dose levels, and the ability to discriminate between Aβ- and Aβ + scans was evaluated using ROC analyses. Finally, intra-reader agreement between the reduced dose and 100% images was evaluated.

RESULTS

At 5% injected dose, change in SUVR was 3.72% and 3.12%, with absolute change in Centiloid 3.35CL and 4.62CL, for [18F]flutemetamol and [18F]florbetaben, respectively. At 12.5% injected dose, percentage change in SUVR and absolute change in Centiloid were < 1.5%. AUCs for discriminating Aβ- from Aβ + scans were high (AUC ≥ 0.94) across dose levels, and visual assessment showed intra-reader agreement of > 80% for both tracers.

CONCLUSION

This proof-of-concept study showed that for both [18F]flutemetamol and [18F]florbetaben, adequate quantitative and qualitative assessments can be obtained at 12.5% of the original injected dose. However, decisions to reduce the injected dose should be made considering the specific clinical or research circumstances.

The quest for multifunctional and dedicated PET instrumentation with irregular geometries

We focus on reviewing state-of-the-art developments of dedicated PET scanners with irregular geometries and the potential of different aspects of multifunctional PET imaging. First, we discuss advances in non-conventional PET detector geometries. Then, we present innovative designs of organ-specific dedicated PET scanners for breast, brain, prostate, and cardiac imaging. We will also review challenges and possible artifacts by image reconstruction algorithms for PET scanners with irregular geometries, such as non-cylindrical and partial angular coverage geometries and how they can be addressed. Then, we attempt to address some open issues about cost/benefits analysis of dedicated PET scanners, how far are the theoretical conceptual designs from the market/clinic, and strategies to reduce fabrication cost without compromising performance.

Differential privacy preserved federated transfer learning for multi-institutional 68Ga-PET image artefact detection and disentanglement

PURPOSE

Image artefacts continue to pose challenges in clinical molecular imaging, resulting in misdiagnoses, additional radiation doses to patients and financial costs. Mismatch and halo artefacts occur frequently in gallium-68 (68Ga)-labelled compounds whole-body PET/CT imaging. Correcting for these artefacts is not straightforward and requires algorithmic developments, given that conventional techniques have failed to address them adequately. In the current study, we employed differential privacy-preserving federated transfer learning (FTL) to manage clinical data sharing and tackle privacy issues for building centre-specific models that detect and correct artefacts present in PET images.

METHODS

Altogether, 1413 patients with 68Ga prostate-specific membrane antigen (PSMA)/DOTA-TATE (TOC) PET/CT scans from 3 countries, including 8 different centres, were enrolled in this study. CT-based attenuation and scatter correction (CT-ASC) was used in all centres for quantitative PET reconstruction. Prior to model training, an experienced nuclear medicine physician reviewed all images to ensure the use of high-quality, artefact-free PET images (421 patients' images). A deep neural network (modified U2Net) was trained on 80% of the artefact-free PET images to utilize centre-based (CeBa), centralized (CeZe) and the proposed differential privacy FTL frameworks. Quantitative analysis was performed in 20% of the clean data (with no artefacts) in each centre. A panel of two nuclear medicine physicians conducted qualitative assessment of image quality, diagnostic confidence and image artefacts in 128 patients with artefacts (256 images for CT-ASC and FTL-ASC).

RESULTS

The three approaches investigated in this study for 68Ga-PET imaging (CeBa, CeZe and FTL) resulted in a mean absolute error (MAE) of 0.42 ± 0.21 (CI 95%: 0.38 to 0.47), 0.32 ± 0.23 (CI 95%: 0.27 to 0.37) and 0.28 ± 0.15 (CI 95%: 0.25 to 0.31), respectively. Statistical analysis using the Wilcoxon test revealed significant differences between the three approaches, with FTL outperforming CeBa and CeZe (p-value < 0.05) in the clean test set. The qualitative assessment demonstrated that FTL-ASC significantly improved image quality and diagnostic confidence and decreased image artefacts, compared to CT-ASC in 68Ga-PET imaging. In addition, mismatch and halo artefacts were successfully detected and disentangled in the chest, abdomen and pelvic regions in 68Ga-PET imaging.

CONCLUSION

The proposed approach benefits from using large datasets from multiple centres while preserving patient privacy. Qualitative assessment by nuclear medicine physicians showed that the proposed model correctly addressed two main challenging artefacts in 68Ga-PET imaging. This technique could be integrated in the clinic for 68Ga-PET imaging artefact detection and disentanglement using multicentric heterogeneous datasets.

Image reconstruction using UNET-transformer network for fast and low-dose PET scans

INTRODUCTION

Low-dose and fast PET imaging (low-count PET) play a significant role in enhancing patient safety, healthcare efficiency, and patient comfort during medical imaging procedures. To achieve high-quality images with low-count PET scans, effective reconstruction models are crucial for denoising and enhancing image quality. The main goal of this paper is to develop an effective and accurate deep learning-based method for reconstructing low-count PET images, which is a challenging problem due to the limited amount of available data and the high level of noise in the acquired images. The proposed method aims to improve the quality of reconstructed PET images while preserving important features, such as edges and small details, by combining the strengths of UNET and Transformer networks.

MATERIAL AND METHODS

The proposed TrUNET-MAPEM model integrates a residual UNET-transformer regularizer into the unrolled maximum a posteriori expectation maximization (MAPEM) algorithm for PET image reconstruction. A loss function based on a combination of structural similarity index (SSIM) and mean squared error (MSE) is utilized to evaluate the accuracy of the reconstructed images. The simulated dataset was generated using the Brainweb phantom, while the real patient dataset was acquired using a Siemens Biograph mMR PET scanner. We also implemented state-of-the-art methods for comparison purposes: OSEM, MAPOSEM, and supervised learning using 3D-UNET network. The reconstructed images are compared to ground truth images using metrics such as peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and relative root mean square error (rRMSE) to quantitatively evaluate the accuracy of the reconstructed images.

RESULTS

Our proposed TrUNET-MAPEM approach was evaluated using both simulated and real patient data. For the patient data, our model achieved an average PSNR of 33.72 dB, an average SSIM of 0.955, and an average rRMSE of 0.39. These results outperformed other methods which had average PSNRs of 36.89 dB, 34.12 dB, and 33.52 db, average SSIMs of 0.944, 0.947, and 0.951, and average rRMSEs of 0.59, 0.49, and 0.42. For the simulated data, our model achieved an average PSNR of 31.23 dB, an average SSIM of 0.95, and an average rRMSE of 0.55. These results also outperformed other state-of-the-art methods, such as OSEM, MAPOSEM, and 3DUNET-MAPEM. The model demonstrates the potential for clinical use by successfully reconstructing smooth images while preserving edges. The comparison with other methods demonstrates the superiority of our approach, as it outperforms all other methods for all three metrics.

CONCLUSION

The proposed TrUNET-MAPEM model presents a significant advancement in the field of low-count PET image reconstruction. The results demonstrate the potential for clinical use, as the model can produce images with reduced noise levels and better edge preservation compared to other reconstruction and post-processing algorithms. The proposed approach may have important clinical applications in the early detection and diagnosis of various diseases.

Artificial Intelligence-Driven Single-Shot PET Image Artifact Detection and Disentanglement: Toward Routine Clinical Image Quality Assurance

PURPOSE

Medical imaging artifacts compromise image quality and quantitative analysis and might confound interpretation and misguide clinical decision-making. The present work envisions and demonstrates a new paradigm PET image Quality Assurance NETwork (PET-QA-NET) in which various image artifacts are detected and disentangled from images without prior knowledge of a standard of reference or ground truth for routine PET image quality assurance.

METHODS

The network was trained and evaluated using training/validation/testing data sets consisting of 669/100/100 artifact-free oncological 18 F-FDG PET/CT images and subsequently fine-tuned and evaluated on 384 (20% for fine-tuning) scans from 8 different PET centers. The developed DL model was quantitatively assessed using various image quality metrics calculated for 22 volumes of interest defined on each scan. In addition, 200 additional 18 F-FDG PET/CT scans (this time with artifacts), generated using both CT-based attenuation and scatter correction (routine PET) and PET-QA-NET, were blindly evaluated by 2 nuclear medicine physicians for the presence of artifacts, diagnostic confidence, image quality, and the number of lesions detected in different body regions.

RESULTS

Across the volumes of interest of 100 patients, SUV MAE values of 0.13 ± 0.04, 0.24 ± 0.1, and 0.21 ± 0.06 were reached for SUV mean , SUV max , and SUV peak , respectively (no statistically significant difference). Qualitative assessment showed a general trend of improved image quality and diagnostic confidence and reduced image artifacts for PET-QA-NET compared with routine CT-based attenuation and scatter correction.

CONCLUSION

We developed a highly effective and reliable quality assurance tool that can be embedded routinely to detect and correct for 18 F-FDG PET image artifacts in clinical setting with notably improved PET image quality and quantitative capabilities.

An active learning approach to train a deep learning algorithm for tumor segmentation from brain MR images

PURPOSE

This study focuses on assessing the performance of active learning techniques to train a brain MRI glioma segmentation model.

METHODS

The publicly available training dataset provided for the 2021 RSNA-ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge was used in this study, consisting of 1251 multi-institutional, multi-parametric MR images. Post-contrast T1, T2, and T2 FLAIR images as well as ground truth manual segmentation were used as input for the model. The data were split into a training set of 1151 cases and testing set of 100 cases, with the testing set remaining constant throughout. Deep convolutional neural network segmentation models were trained using the NiftyNet platform. To test the viability of active learning in training a segmentation model, an initial reference model was trained using all 1151 training cases followed by two additional models using only 575 cases and 100 cases. The resulting predicted segmentations of these two additional models on the remaining training cases were then addended to the training dataset for additional training.

RESULTS

It was demonstrated that an active learning approach for manual segmentation can lead to comparable model performance for segmentation of brain gliomas (0.906 reference Dice score vs 0.868 active learning Dice score) while only requiring manual annotation for 28.6% of the data.

CONCLUSION

The active learning approach when applied to model training can drastically reduce the time and labor spent on preparation of ground truth training data.

CRITICAL RELEVANCE STATEMENT

Active learning concepts were applied to a deep learning-assisted segmentation of brain gliomas from MR images to assess their viability in reducing the required amount of manually annotated ground truth data in model training.

KEY POINTS

• This study focuses on assessing the performance of active learning techniques to train a brain MRI glioma segmentation model. • The active learning approach for manual segmentation can lead to comparable model performance for segmentation of brain gliomas. • Active learning when applied to model training can drastically reduce the time and labor spent on preparation of ground truth training data.

Reduction of scan duration and radiation dose in cerebral CT perfusion imaging of acute stroke using a recurrent neural network

Objective. Cerebral CT perfusion (CTP) imaging is most commonly used to diagnose acute ischaemic stroke and support treatment decisions. Shortening CTP scan duration is desirable to reduce the accumulated radiation dose and the risk of patient head movement. In this study, we present a novel application of a stochastic adversarial video prediction approach to reduce CTP imaging acquisition time.Approach. A variational autoencoder and generative adversarial network (VAE-GAN) were implemented in a recurrent framework in three scenarios: to predict the last 8 (24 s), 13 (31.5 s) and 18 (39 s) image frames of the CTP acquisition from the first 25 (36 s), 20 (28.5 s) and 15 (21 s) acquired frames, respectively. The model was trained using 65 stroke cases and tested on 10 unseen cases. Predicted frames were assessed against ground-truth in terms of image quality and haemodynamic maps, bolus shape characteristics and volumetric analysis of lesions.Main results. In all three prediction scenarios, the mean percentage error between the area, full-width-at-half-maximum and maximum enhancement of the predicted and ground-truth bolus curve was less than 4 ± 4%. The best peak signal-to-noise ratio and structural similarity of predicted haemodynamic maps was obtained for cerebral blood volume followed (in order) by cerebral blood flow, mean transit time and time to peak. For the 3 prediction scenarios, average volumetric error of the lesion was overestimated by 7%-15%, 11%-28% and 7%-22% for the infarct, penumbra and hypo-perfused regions, respectively, and the corresponding spatial agreement for these regions was 67%-76%, 76%-86% and 83%-92%.Significance. This study suggests that a recurrent VAE-GAN could potentially be used to predict a portion of CTP frames from truncated acquisitions, preserving the majority of clinical content in the images, and potentially reducing the scan duration and radiation dose simultaneously by 65% and 54.5%, respectively.

A cycle-consistent adversarial network for brain PET partial volume correction without prior anatomical information

PURPOSE

Partial volume effect (PVE) is a consequence of the limited spatial resolution of PET scanners. PVE can cause the intensity values of a particular voxel to be underestimated or overestimated due to the effect of surrounding tracer uptake. We propose a novel partial volume correction (PVC) technique to overcome the adverse effects of PVE on PET images.

METHODS

Two hundred and twelve clinical brain PET scans, including 50 18F-Fluorodeoxyglucose (18F-FDG), 50 18F-Flortaucipir, 36 18F-Flutemetamol, and 76 18F-FluoroDOPA, and their corresponding T1-weighted MR images were enrolled in this study. The Iterative Yang technique was used for PVC as a reference or surrogate of the ground truth for evaluation. A cycle-consistent adversarial network (CycleGAN) was trained to directly map non-PVC PET images to PVC PET images. Quantitative analysis using various metrics, including structural similarity index (SSIM), root mean squared error (RMSE), and peak signal-to-noise ratio (PSNR), was performed. Furthermore, voxel-wise and region-wise-based correlations of activity concentration between the predicted and reference images were evaluated through joint histogram and Bland and Altman analysis. In addition, radiomic analysis was performed by calculating 20 radiomic features within 83 brain regions. Finally, a voxel-wise two-sample t-test was used to compare the predicted PVC PET images with reference PVC images for each radiotracer.

RESULTS

The Bland and Altman analysis showed the largest and smallest variance for 18F-FDG (95% CI: - 0.29, + 0.33 SUV, mean = 0.02 SUV) and 18F-Flutemetamol (95% CI: - 0.26, + 0.24 SUV, mean =  - 0.01 SUV), respectively. The PSNR was lowest (29.64 ± 1.13 dB) for 18F-FDG and highest (36.01 ± 3.26 dB) for 18F-Flutemetamol. The smallest and largest SSIM were achieved for 18F-FDG (0.93 ± 0.01) and 18F-Flutemetamol (0.97 ± 0.01), respectively. The average relative error for the kurtosis radiomic feature was 3.32%, 9.39%, 4.17%, and 4.55%, while it was 4.74%, 8.80%, 7.27%, and 6.81% for NGLDM_contrast feature for 18F-Flutemetamol, 18F-FluoroDOPA, 18F-FDG, and 18F-Flortaucipir, respectively.

CONCLUSION

An end-to-end CycleGAN PVC method was developed and evaluated. Our model generates PVC images from the original non-PVC PET images without requiring additional anatomical information, such as MRI or CT. Our model eliminates the need for accurate registration or segmentation or PET scanner system response characterization. In addition, no assumptions regarding anatomical structure size, homogeneity, boundary, or background level are required.

Decentralized collaborative multi-institutional PET attenuation and scatter correction using federated deep learning

PURPOSE

Attenuation correction and scatter compensation (AC/SC) are two main steps toward quantitative PET imaging, which remain challenging in PET-only and PET/MRI systems. These can be effectively tackled via deep learning (DL) methods. However, trustworthy, and generalizable DL models commonly require well-curated, heterogeneous, and large datasets from multiple clinical centers. At the same time, owing to legal/ethical issues and privacy concerns, forming a large collective, centralized dataset poses significant challenges. In this work, we aimed to develop a DL-based model in a multicenter setting without direct sharing of data using federated learning (FL) for AC/SC of PET images.

METHODS

Non-attenuation/scatter corrected and CT-based attenuation/scatter corrected (CT-ASC) ¹⁸F-FDG PET images of 300 patients were enrolled in this study. The dataset consisted of 6 different centers, each with 50 patients, with scanner, image acquisition, and reconstruction protocols varying across the centers. CT-based ASC PET images served as the standard reference. All images were reviewed to include high-quality and artifact-free PET images. Both corrected and uncorrected PET images were converted to standardized uptake values (SUVs). We used a modified nested U-Net utilizing residual U-block in a U-shape architecture. We evaluated two FL models, namely sequential (FL-SQ) and parallel (FL-PL) and compared their performance with the baseline centralized (CZ) learning model wherein the data were pooled to one server, as well as center-based (CB) models where for each center the model was built and evaluated separately. Data from each center were divided to contribute to training (30 patients), validation (10 patients), and test sets (10 patients). Final evaluations and reports were performed on 60 patients (10 patients from each center).

RESULTS

In terms of percent SUV absolute relative error (ARE%), both FL-SQ (CI:12.21-14.81%) and FL-PL (CI:11.82-13.84%) models demonstrated excellent agreement with the centralized framework (CI:10.32-12.00%), while FL-based algorithms improved model performance by over 11% compared to CB training strategy (CI: 22.34-26.10%). Furthermore, the Mann-Whitney test between different strategies revealed no significant differences between CZ and FL-based algorithms (p-value > 0.05) in center-categorized mode. At the same time, a significant difference was observed between the different training approaches on the overall dataset (p-value < 0.05). In addition, voxel-wise comparison, with respect to reference CT-ASC, exhibited similar performance for images predicted by CZ (R² = 0.94), FL-SQ (R² = 0.93), and FL-PL (R² = 0.92), while CB model achieved a far lower coefficient of determination (R² = 0.74). Despite the strong correlations between CZ and FL-based methods compared to reference CT-ASC, a slight underestimation of predicted voxel values was observed.

CONCLUSION

Deep learning-based models provide promising results toward quantitative PET image reconstruction. Specifically, we developed two FL models and compared their performance with center-based and centralized models. The proposed FL-based models achieved higher performance compared to center-based models, comparable with centralized models. Our work provided strong empirical evidence that the FL framework can fully benefit from the generalizability and robustness of DL models used for AC/SC in PET, while obviating the need for the direct sharing of datasets between clinical imaging centers.

Deep-TOF-PET: Deep learning-guided generation of time-of-flight from non-TOF brain PET images in the image and projection domains

We aim to synthesize brain time‐of‐flight (TOF) PET images/sinograms from their corresponding non‐TOF information in the image space (IS) and sinogram space (SS) to increase the signal‐to‐noise ratio (SNR) and contrast of abnormalities, and decrease the bias in tracer uptake quantification. One hundred forty clinical brain ¹⁸F‐FDG PET/CT scans were collected to generate TOF and non‐TOF sinograms. The TOF sinograms were split into seven time bins (0, ±1, ±2, ±3). The predicted TOF sinogram was reconstructed and the performance of both models (IS and SS) compared with reference TOF and non‐TOF. Wide‐ranging quantitative and statistical analysis metrics, including structural similarity index metric (SSIM), root mean square error (RMSE), as well as 28 radiomic features for 83 brain regions were extracted to evaluate the performance of the CycleGAN model. SSIM and RMSE of 0.99 ± 0.03, 0.98 ± 0.02 and 0.12 ± 0.09, 0.16 ± 0.04 were achieved for the generated TOF‐PET images in IS and SS, respectively. They were 0.97 ± 0.03 and 0.22 ± 0.12, respectively, for non‐TOF‐PET images. The Bland & Altman analysis revealed that the lowest tracer uptake value bias (−0.02%) and minimum variance (95% CI: −0.17%, +0.21%) were achieved for TOF‐PET images generated in IS. For malignant lesions, the contrast in the test dataset was enhanced from 3.22 ± 2.51 for non‐TOF to 3.34 ± 0.41 and 3.65 ± 3.10 for TOF PET in SS and IS, respectively. The implemented CycleGAN is capable of generating TOF from non‐TOF PET images to achieve better image quality.

High-dimensional multinomial multiclass severity scoring of COVID-19 pneumonia using CT radiomics features and machine learning algorithms

We aimed to construct a prediction model based on computed tomography (CT) radiomics features to classify COVID-19 patients into severe-, moderate-, mild-, and non-pneumonic. A total of 1110 patients were studied from a publicly available dataset with 4-class severity scoring performed by a radiologist (based on CT images and clinical features). The entire lungs were segmented and followed by resizing, bin discretization and radiomic features extraction. We utilized two feature selection algorithms, namely bagging random forest (BRF) and multivariate adaptive regression splines (MARS), each coupled to a classifier, namely multinomial logistic regression (MLR), to construct multiclass classification models. The dataset was divided into 50% (555 samples), 20% (223 samples), and 30% (332 samples) for training, validation, and untouched test datasets, respectively. Subsequently, nested cross-validation was performed on train/validation to select the features and tune the models. All predictive power indices were reported based on the testing set. The performance of multi-class models was assessed using precision, recall, F1-score, and accuracy based on the 4 × 4 confusion matrices. In addition, the areas under the receiver operating characteristic curves (AUCs) for multi-class classifications were calculated and compared for both models. Using BRF, 23 radiomic features were selected, 11 from first-order, 9 from GLCM, 1 GLRLM, 1 from GLDM, and 1 from shape. Ten features were selected using the MARS algorithm, namely 3 from first-order, 1 from GLDM, 1 from GLRLM, 1 from GLSZM, 1 from shape, and 3 from GLCM features. The mean absolute deviation, skewness, and variance from first-order and flatness from shape, and cluster prominence from GLCM features and Gray Level Non Uniformity Normalize from GLRLM were selected by both BRF and MARS algorithms. All selected features by BRF or MARS were significantly associated with four-class outcomes as assessed within MLR (All p values < 0.05). BRF + MLR and MARS + MLR resulted in pseudo-R² prediction performances of 0.305 and 0.253, respectively. Meanwhile, there was a significant difference between the feature selection models when using a likelihood ratio test (p value = 0.046). Based on confusion matrices for BRF + MLR and MARS + MLR algorithms, the precision was 0.856 and 0.728, the recall was 0.852 and 0.722, whereas the accuracy was 0.921 and 0.861, respectively. AUCs (95% CI) for multi-class classification were 0.846 (0.805-0.887) and 0.807 (0.752-0.861) for BRF + MLR and MARS + MLR algorithms, respectively. Our models based on the utilization of radiomic features, coupled with machine learning were able to accurately classify patients according to the severity of pneumonia, thus highlighting the potential of this emerging paradigm in the prognostication and management of COVID-19 patients.

Active-PET: a multifunctional PET scanner with dynamic gantry size featuring high-resolution and high-sensitivity imaging: a Monte Carlo simulation study

Organ-specific PET scanners have been developed to provide both high spatial resolution and sensitivity, although the deployment of several dedicated PET scanners at the same center is costly and space-consuming. Active-PET is a multifunctional PET scanner design exploiting the advantages of two different types of detector modules and mechanical arms mechanisms enabling repositioning of the detectors to allow the implementation of different geometries/configurations. Active-PET can be used for different applications, including brain, axilla, breast, prostate, whole-body, preclinical and pediatrics imaging, cell tracking, and image guidance for therapy. Monte Carlo techniques were used to simulate a PET scanner with two sets of high resolution and high sensitivity pixelated Lutetium Oxyorthoscilicate (LSO(Ce)) detector blocks (24 for each group, overall 48 detector modules for each ring), one with large pixel size (4 × 4 mm2) and crystal thickness (20 mm), and another one with small pixel size (2 × 2 mm2) and thickness (10 mm). Each row of detector modules is connected to a linear motor that can displace the detectors forward and backward along the radial axis to achieve variable gantry diameter in order to image the target subject at the optimal/desired resolution and/or sensitivity. At the center of the field-of-view, the highest sensitivity (15.98 kcps MBq−1) was achieved by the scanner with a small gantry and high-sensitivity detectors while the best spatial resolution was obtained by the scanner with a small gantry and high-resolution detectors (2.2 mm, 2.3 mm, 2.5 mm FWHM for tangential, radial, and axial, respectively). The configuration with large-bore (combination of high-resolution and high-sensitivity detectors) achieved better performance and provided higher image quality compared to the Biograph mCT as reflected by the 3D Hoffman brain phantom simulation study. We introduced the concept of a non-static PET scanner capable of switching between large and small field-of-view as well as high-resolution and high-sensitivity imaging.

Robust-Deep: A Method for Increasing Brain Imaging Datasets to Improve Deep Learning Models' Performance and Robustness

A small dataset commonly affects generalization, robustness, and overall performance of deep neural networks (DNNs) in medical imaging research. Since gathering large clinical databases is always difficult, we proposed an analytical method for producing a large realistic/diverse dataset. Clinical brain PET/CT/MR images including full-dose (FD), low-dose (LD) corresponding to only 5 % of events acquired in the FD scan, non-attenuated correction (NAC) and CT-based measured attenuation correction (MAC) PET images, CT images and T1 and T2 MR sequences of 35 patients were included. All images were registered to the Montreal Neurological Institute (MNI) template. Laplacian blending was used to make a natural presentation using information in the frequency domain of images from two separate patients, as well as the blending mask. This classical technique from the computer vision and image processing communities is still widely used and unlike modern DNNs, does not require the availability of training data. A modified ResNet DNN was implemented to evaluate four image-to-image translation tasks, including LD to FD, LD+MR to FD, NAC to MAC, and MRI to CT, with and without using the synthesized images. Quantitative analysis using established metrics, including the peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and joint histogram analysis was performed for quantitative evaluation. The quantitative comparison between the registered small dataset containing 35 patients and the large dataset containing 350 synthesized plus 35 real dataset demonstrated improvement of the RMSE and SSIM by 29% and 8% for LD to FD, 40% and 7% for LD+MRI to FD, 16% and 8% for NAC to MAC, and 24% and 11% for MRI to CT mapping task, respectively. The qualitative/quantitative analysis demonstrated that the proposed model improved the performance of all four DNN models through producing images of higher quality and lower quantitative bias and variance compared to reference images.

COVID-19 prognostic modeling using CT radiomic features and machine learning algorithms: Analysis of a multi-institutional dataset of 14,339 patients

BACKGROUND

We aimed to analyze the prognostic power of CT-based radiomics models using data of 14,339 COVID-19 patients.

METHODS

Whole lung segmentations were performed automatically using a deep learning-based model to extract 107 intensity and texture radiomics features. We used four feature selection algorithms and seven classifiers. We evaluated the models using ten different splitting and cross-validation strategies, including non-harmonized and ComBat-harmonized datasets. The sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were reported.

RESULTS

In the test dataset (4,301) consisting of CT and/or RT-PCR positive cases, AUC, sensitivity, and specificity of 0.83 ± 0.01 (CI95%: 0.81-0.85), 0.81, and 0.72, respectively, were obtained by ANOVA feature selector + Random Forest (RF) classifier. Similar results were achieved in RT-PCR-only positive test sets (3,644). In ComBat harmonized dataset, Relief feature selector + RF classifier resulted in the highest performance of AUC, reaching 0.83 ± 0.01 (CI95%: 0.81-0.85), with a sensitivity and specificity of 0.77 and 0.74, respectively. ComBat harmonization did not depict statistically significant improvement compared to a non-harmonized dataset. In leave-one-center-out, the combination of ANOVA feature selector and RF classifier resulted in the highest performance.

CONCLUSION

Lung CT radiomics features can be used for robust prognostic modeling of COVID-19. The predictive power of the proposed CT radiomics model is more reliable when using a large multicentric heterogeneous dataset, and may be used prospectively in clinical setting to manage COVID-19 patients.

COLI-Net: Deep learning-assisted fully automated COVID-19 lung and infection pneumonia lesion detection and segmentation from chest computed tomography images

We present a deep learning (DL)-based automated whole lung and COVID-19 pneumonia infectious lesions (COLI-Net) detection and segmentation from chest computed tomography (CT) images. This multicenter/multiscanner study involved 2368 (347'259 2D slices) and 190 (17 341 2D slices) volumetric CT exams along with their corresponding manual segmentation of lungs and lesions, respectively. All images were cropped, resized, and the intensity values clipped and normalized. A residual network with non-square Dice loss function built upon TensorFlow was employed. The accuracy of lung and COVID-19 lesions segmentation was evaluated on an external reverse transcription-polymerase chain reaction positive COVID-19 dataset (7'333 2D slices) collected at five different centers. To evaluate the segmentation performance, we calculated different quantitative metrics, including radiomic features. The mean Dice coefficients were 0.98 ± 0.011 (95% CI, 0.98-0.99) and 0.91 ± 0.038 (95% CI, 0.90-0.91) for lung and lesions segmentation, respectively. The mean relative Hounsfield unit differences were 0.03 ± 0.84% (95% CI, -0.12 to 0.18) and -0.18 ± 3.4% (95% CI, -0.8 to 0.44) for the lung and lesions, respectively. The relative volume difference for lung and lesions were 0.38 ± 1.2% (95% CI, 0.16-0.59) and 0.81 ± 6.6% (95% CI, -0.39 to 2), respectively. Most radiomic features had a mean relative error less than 5% with the highest mean relative error achieved for the lung for the range first-order feature (-6.95%) and least axis length shape feature (8.68%) for lesions. We developed an automated DL-guided three-dimensional whole lung and infected regions segmentation in COVID-19 patients to provide fast, consistent, robust, and human error immune framework for lung and pneumonia lesion detection and quantification.

Deep learning-based fully automated Z-axis coverage range definition from scout scans to eliminate overscanning in chest CT imaging

BACKGROUND

Despite the prevalence of chest CT in the clinic, concerns about unoptimized protocols delivering high radiation doses to patients still remain. This study aimed to assess the additional radiation dose associated with overscanning in chest CT and to develop an automated deep learning-assisted scan range selection technique to reduce radiation dose to patients.

RESULTS

A significant overscanning range (31 ± 24) mm was observed in clinical setting for over 95% of the cases. The average Dice coefficient for lung segmentation was 0.96 and 0.97 for anterior-posterior (AP) and lateral projections, respectively. By considering the exact lung coverage as the ground truth, and AP and lateral projections as input, The DL-based approach resulted in errors of 0.08 ± 1.46 and - 1.5 ± 4.1 mm in superior and inferior directions, respectively. In contrast, the error on external scout views was - 0.7 ± 4.08 and 0.01 ± 14.97 mm for superior and inferior directions, respectively.The ED reduction achieved by automated scan range selection was 21% in the test group. The evaluation of a large multi-centric chest CT dataset revealed unnecessary ED of more than 2 mSv per scan and 67% increase in the thyroid absorbed dose.

CONCLUSION

The proposed DL-based solution outperformed previous automatic methods with acceptable accuracy, even in complicated and challenging cases. The generizability of the model was demonstrated by fine-tuning the model on AP scout views and achieving acceptable results. The method can reduce the unoptimized dose to patients by exclunding unnecessary organs from field of view.

DeepTOFSino: A deep learning model for synthesizing full-dose time-of-flight bin sinograms from their corresponding low-dose sinograms

PURPOSE

Reducing the injected activity and/or the scanning time is a desirable goal to minimize radiation exposure and maximize patients' comfort. To achieve this goal, we developed a deep neural network (DNN) model for synthesizing full-dose (FD) time-of-flight (TOF) bin sinograms from their corresponding fast/low-dose (LD) TOF bin sinograms.

METHODS

Clinical brain PET/CT raw data of 140 normal and abnormal patients were employed to create LD and FD TOF bin sinograms. The LD TOF sinograms were created through 5% undersampling of FD list-mode PET data. The TOF sinograms were split into seven time bins (0, ±1, ±2, ±3). Residual network (ResNet) algorithms were trained separately to generate FD bins from LD bins. An extra ResNet model was trained to synthesize FD images from LD images to compare the performance of DNN in sinogram space (SS) vs implementation in image space (IS). Comprehensive quantitative and statistical analysis was performed to assess the performance of the proposed model using established quantitative metrics, including the peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM) region-wise standardized uptake value (SUV) bias and statistical analysis for 83 brain regions.

RESULTS

SSIM and PSNR values of 0.97 ± 0.01, 0.98 ± 0.01 and 33.70 ± 0.32, 39.36 ± 0.21 were obtained for IS and SS, respectively, compared to 0.86 ± 0.02and 31.12 ± 0.22 for reference LD images. The absolute average SUV bias was 0.96 ± 0.95% and 1.40 ± 0.72% for SS and IS implementations, respectively. The joint histogram analysis revealed the lowest mean square error (MSE) and highest correlation (R² = 0.99, MSE = 0.019) was achieved by SS compared to IS (R² = 0.97, MSE= 0.028). The Bland & Altman analysis showed that the lowest SUV bias (-0.4%) and minimum variance (95% CI: -2.6%, +1.9%) were achieved by SS images. The voxel-wise t-test analysis revealed the presence of voxels with statistically significantly lower values in LD, IS, and SS images compared to FD images respectively.

CONCLUSION

The results demonstrated that images reconstructed from the predicted TOF FD sinograms using the SS approach led to higher image quality and lower bias compared to images predicted from LD images.

Fully Automated Gross Tumor Volume Delineation From PET in Head and Neck Cancer Using Deep Learning Algorithms

PURPOSE

The availability of automated, accurate, and robust gross tumor volume (GTV) segmentation algorithms is critical for the management of head and neck cancer (HNC) patients. In this work, we evaluated 3 state-of-the-art deep learning algorithms combined with 8 different loss functions for PET image segmentation using a comprehensive training set and evaluated its performance on an external validation set of HNC patients.

PATIENTS AND METHODS

18F-FDG PET/CT images of 470 patients presenting with HNC on which manually defined GTVs serving as standard of reference were used for training (340 patients), evaluation (30 patients), and testing (100 patients from different centers) of these algorithms. PET image intensity was converted to SUVs and normalized in the range (0-1) using the SUVmax of the whole data set. PET images were cropped to 12 × 12 × 12 cm3 subvolumes using isotropic voxel spacing of 3 × 3 × 3 mm3 containing the whole tumor and neighboring background including lymph nodes. We used different approaches for data augmentation, including rotation (-15 degrees, +15 degrees), scaling (-20%, 20%), random flipping (3 axes), and elastic deformation (sigma = 1 and proportion to deform = 0.7) to increase the number of training sets. Three state-of-the-art networks, including Dense-VNet, NN-UNet, and Res-Net, with 8 different loss functions, including Dice, generalized Wasserstein Dice loss, Dice plus XEnt loss, generalized Dice loss, cross-entropy, sensitivity-specificity, and Tversky, were used. Overall, 28 different networks were built. Standard image segmentation metrics, including Dice similarity, image-derived PET metrics, first-order, and shape radiomic features, were used for performance assessment of these algorithms.

RESULTS

The best results in terms of Dice coefficient (mean ± SD) were achieved by cross-entropy for Res-Net (0.86 ± 0.05; 95% confidence interval [CI], 0.85-0.87), Dense-VNet (0.85 ± 0.058; 95% CI, 0.84-0.86), and Dice plus XEnt for NN-UNet (0.87 ± 0.05; 95% CI, 0.86-0.88). The difference between the 3 networks was not statistically significant (P > 0.05). The percent relative error (RE%) of SUVmax quantification was less than 5% in networks with a Dice coefficient more than 0.84, whereas a lower RE% (0.41%) was achieved by Res-Net with cross-entropy loss. For maximum 3-dimensional diameter and sphericity shape features, all networks achieved a RE ≤ 5% and ≤10%, respectively, reflecting a small variability.

CONCLUSIONS

Deep learning algorithms exhibited promising performance for automated GTV delineation on HNC PET images. Different loss functions performed competitively when using different networks and cross-entropy for Res-Net, Dense-VNet, and Dice plus XEnt for NN-UNet emerged as reliable networks for GTV delineation. Caution should be exercised for clinical deployment owing to the occurrence of outliers in deep learning-based algorithms.

Deep learning-assisted ultra-fast/low-dose whole-body PET/CT imaging

PURPOSE

Tendency is to moderate the injected activity and/or reduce acquisition time in PET examinations to minimize potential radiation hazards and increase patient comfort. This work aims to assess the performance of regular full-dose (FD) synthesis from fast/low-dose (LD) whole-body (WB) PET images using deep learning techniques.

METHODS

Instead of using synthetic LD scans, two separate clinical WB 18F-Fluorodeoxyglucose (18F-FDG) PET/CT studies of 100 patients were acquired: one regular FD (~ 27 min) and one fast or LD (~ 3 min) consisting of 1/8th of the standard acquisition time. A modified cycle-consistent generative adversarial network (CycleGAN) and residual neural network (ResNET) models, denoted as CGAN and RNET, respectively, were implemented to predict FD PET images. The quality of the predicted PET images was assessed by two nuclear medicine physicians. Moreover, the diagnostic quality of the predicted PET images was evaluated using a pass/fail scheme for lesion detectability task. Quantitative analysis using established metrics including standardized uptake value (SUV) bias was performed for the liver, left/right lung, brain, and 400 malignant lesions from the test and evaluation datasets.

RESULTS

CGAN scored 4.92 and 3.88 (out of 5) (adequate to good) for brain and neck + trunk, respectively. The average SUV bias calculated over normal tissues was 3.39 ± 0.71% and - 3.83 ± 1.25% for CGAN and RNET, respectively. Bland-Altman analysis reported the lowest SUV bias (0.01%) and 95% confidence interval of - 0.36, + 0.47 for CGAN compared with the reference FD images for malignant lesions.

CONCLUSION

CycleGAN is able to synthesize clinical FD WB PET images from LD images with 1/8th of standard injected activity or acquisition time. The predicted FD images present almost similar performance in terms of lesion detectability, qualitative scores, and quantification bias and variance.

Fast dynamic brain PET imaging using stochastic variational prediction for recurrent frame generation

Purpose

We assess the performance of a recurrent frame generation algorithm for prediction of late frames from initial frames in dynamic brain PET imaging.

Methods

Clinical dynamic ¹⁸F‐DOPA brain PET/CT studies of 46 subjects with ten folds cross‐validation were retrospectively employed. A novel stochastic adversarial video prediction model was implemented to predict the last 13 frames (25–90 minutes) from the initial 13 frames (0–25 minutes). The quantitative analysis of the predicted dynamic PET frames was performed for the test and validation dataset using established metrics.

Results

The predicted dynamic images demonstrated that the model is capable of predicting the trend of change in time‐varying tracer biodistribution. The Bland‐Altman plots reported the lowest tracer uptake bias (−0.04) for the putamen region and the smallest variance (95% CI: −0.38, +0.14) for the cerebellum. The region‐wise Patlak graphical analysis in the caudate and putamen regions for eight subjects from the test and validation dataset showed that the average bias for Ki and distribution volume was 4.3%, 5.1% and 4.4%, 4.2%, (P‐value <0.05), respectively.

Conclusion

We have developed a novel deep learning approach for fast dynamic brain PET imaging capable of generating the last 65 minutes time frames from the initial 25 minutes frames, thus enabling significant reduction in scanning time.

Ultra-low-dose chest CT imaging of COVID-19 patients using a deep residual neural network

OBJECTIVES

The current study aimed to design an ultra-low-dose CT examination protocol using a deep learning approach suitable for clinical diagnosis of COVID-19 patients.

METHODS

In this study, 800, 170, and 171 pairs of ultra-low-dose and full-dose CT images were used as input/output as training, test, and external validation set, respectively, to implement the full-dose prediction technique. A residual convolutional neural network was applied to generate full-dose from ultra-low-dose CT images. The quality of predicted CT images was assessed using root mean square error (RMSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). Scores ranging from 1 to 5 were assigned reflecting subjective assessment of image quality and related COVID-19 features, including ground glass opacities (GGO), crazy paving (CP), consolidation (CS), nodular infiltrates (NI), bronchovascular thickening (BVT), and pleural effusion (PE).

RESULTS

The radiation dose in terms of CT dose index (CTDIvol) was reduced by up to 89%. The RMSE decreased from 0.16 ± 0.05 to 0.09 ± 0.02 and from 0.16 ± 0.06 to 0.08 ± 0.02 for the predicted compared with ultra-low-dose CT images in the test and external validation set, respectively. The overall scoring assigned by radiologists showed an acceptance rate of 4.72 ± 0.57 out of 5 for reference full-dose CT images, while ultra-low-dose CT images rated 2.78 ± 0.9. The predicted CT images using the deep learning algorithm achieved a score of 4.42 ± 0.8.

CONCLUSIONS

The results demonstrated that the deep learning algorithm is capable of predicting standard full-dose CT images with acceptable quality for the clinical diagnosis of COVID-19 positive patients with substantial radiation dose reduction.

KEY POINTS

• Ultra-low-dose CT imaging of COVID-19 patients would result in the loss of critical information about lesion types, which could potentially affect clinical diagnosis. • Deep learning-based prediction of full-dose from ultra-low-dose CT images for the diagnosis of COVID-19 could reduce the radiation dose by up to 89%. • Deep learning algorithms failed to recover the correct lesion structure/density for a number of patients considered outliers, and as such, further research and development is warranted to address these limitations.

Polaroid-PET: a PET scanner with detectors fitted with Polaroid for filtering unpolarized optical photons-a Monte Carlo simulation study

We propose and evaluate the performance of an improved preclinical positron emission tomography (PET) scanner design, referred to as Polaroid-PET, consisting of a detector equipped with a layer of horizontal Polaroid to filter scintillation photons with vertical polarization. This makes it possible to improve the spatial resolution of PET scanners based on monolithic crystals. First, a detector module based on a lutetium-yttrium orthosilicate monolithic crystal with 10 mm thickness and silicon photomultipliers (SiPMs) was implemented in the GEANT4 Monte Carlo toolkit. Subsequently, a layer of Polaroid was inserted between the crystal and the SiPMs. Two preclinical PET scanners based on ten detector modules with and without Polaroid were simulated. The performance of the proposed detector modules and corresponding PET scanner for the two configurations (with and without Polaroid) was assessed using standard performance parameters, including spatial resolution, sensitivity, optical photon ratio detected for positioning, and image quality. The detector module fitted with Polaroid led to higher spatial resolution (1.05 mm FWHM) in comparison with a detector without Polaroid (1.30 mm FHWM) for a point source located at the center of the detector module. From 100% of optical photons produced in the scintillator crystal, 65% and 66% were used for positioning in the detectors without and with Polaroid, respectively. Polaroid-PET resulted in higher axial spatial resolution (0.83 mm FWHM) compared to the scanner without Polaroid (1.01 mm FWHM) for a point source at the center of the field of view (CFOV). The absolute sensitivity at the CFOV was 4.37% and 4.31% for regular and Polaroid-PET, respectively. Planar images of a grid phantom demonstrated the potential of the detector with a Polaroid in distinguishing point sources located at close distances. Our results indicated that Polaroid-PET may improve spatial resolution by filtering the reflected optical photons according to their polarization state, while retaining the high sensitivity expected with monolithic crystal detector blocks.

Projection Space Implementation of Deep Learning-Guided Low-Dose Brain PET Imaging Improves Performance over Implementation in Image Space

Our purpose was to assess the performance of full-dose (FD) PET image synthesis in both image and sinogram space from low-dose (LD) PET images and sinograms without sacrificing diagnostic quality using deep learning techniques. Methods: Clinical brain PET/CT studies of 140 patients were retrospectively used for LD-to-FD PET conversion. Five percent of the events were randomly selected from the FD list-mode PET data to simulate a realistic LD acquisition. A modified 3-dimensional U-Net model was implemented to predict FD sinograms in the projection space (PSS) and FD images in image space (PIS) from their corresponding LD sinograms and images, respectively. The quality of the predicted PET images was assessed by 2 nuclear medicine specialists using a 5-point grading scheme. Quantitative analysis using established metrics including the peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), regionwise SUV bias, and first-, second- and high-order texture radiomic features in 83 brain regions for the test and evaluation datasets was also performed. Results: All PSS images were scored 4 or higher (good to excellent) by the nuclear medicine specialists. PSNR and SSIM values of 0.96 ± 0.03 and 0.97 ± 0.02, respectively, were obtained for PIS, and values of 31.70 ± 0.75 and 37.30 ± 0.71, respectively, were obtained for PSS. The average SUV bias calculated over all brain regions was 0.24% ± 0.96% and 1.05% ± 1.44% for PSS and PIS, respectively. The Bland-Altman plots reported the lowest SUV bias (0.02) and variance (95% confidence interval, -0.92 to +0.84) for PSS, compared with the reference FD images. The relative error of the homogeneity radiomic feature belonging to the gray-level cooccurrence matrix category was -1.07 ± 1.77 and 0.28 ± 1.4 for PIS and PSS, respectively. Conclusion: The qualitative assessment and quantitative analysis demonstrated that the FD PET PSS led to superior performance, resulting in higher image quality and lower SUV bias and variance than for FD PET PIS.