Deep learning for [18F]fluorodeoxyglucose-PET-CT classification in patients with lymphoma: a dual-centre retrospective analysis.

BACKGROUND: The rising global cancer burden has led to an increasing demand for imaging tests such as [18F]fluorodeoxyglucose ([18F]FDG)-PET-CT. To aid imaging specialists in dealing with high scan volumes, we aimed to train a deep learning artificial intelligence algorithm to classify [18F]FDG-PET-CT scans of patients with lymphoma with or without hypermetabolic tumour sites. METHODS: In this retrospective analysis we collected 16 583 [18F]FDG-PET-CTs of 5072 patients with lymphoma who had undergone PET-CT before or after treatment at the Memorial Sloa Kettering Cancer Center, New York, NY, USA. Using maximum intensity projection (MIP), three dimensional (3D) PET, and 3D CT data, our ResNet34-based deep learning model (Lymphoma Artificial Reader System [LARS]) for [18F]FDG-PET-CT binary classification (Deauville 1-3 vs 4-5), was trained on 80% of the dataset, and tested on 20% of this dataset. For external testing, 1000 [18F]FDG-PET-CTs were obtained from a second centre (Medical University of Vienna, Vienna, Austria). Seven model variants were evaluated, including MIP-based LARS-avg (optimised for accuracy) and LARS-max (optimised for sensitivity), and 3D PET-CT-based LARS-ptct. Following expert curation, areas under the curve (AUCs), accuracies, sensitivities, and specificities were calculated. FINDINGS: In the internal test cohort (3325 PET-CTs, 1012 patients), LARS-avg achieved an AUC of 0·949 (95% CI 0·942-0·956), accuracy of 0·890 (0·879-0·901), sensitivity of 0·868 (0·851-0·885), and specificity of 0·913 (0·899-0·925); LARS-max achieved an AUC of 0·949 (0·942-0·956), accuracy of 0·868 (0·858-0·879), sensitivity of 0·909 (0·896-0·924), and specificity of 0·826 (0·808-0·843); and LARS-ptct achieved an AUC of 0·939 (0·930-0·948), accuracy of 0·875 (0·864-0·887), sensitivity of 0·836 (0·817-0·855), and specificity of 0·915 (0·901-0·927). In the external test cohort (1000 PET-CTs, 503 patients), LARS-avg achieved an AUC of 0·953 (0·938-0·966), accuracy of 0·907 (0·888-0·925), sensitivity of 0·874 (0·843-0·904), and specificity of 0·949 (0·921-0·960); LARS-max achieved an AUC of 0·952 (0·937-0·965), accuracy of 0·898 (0·878-0·916), sensitivity of 0·899 (0·871-0·926), and specificity of 0·897 (0·871-0·922); and LARS-ptct achieved an AUC of 0·932 (0·915-0·948), accuracy of 0·870 (0·850-0·891), sensitivity of 0·827 (0·793-0·863), and specificity of 0·913 (0·889-0·937). INTERPRETATION: Deep learning accurately distinguishes between [18F]FDG-PET-CT scans of lymphoma patients with and without hypermetabolic tumour sites. Deep learning might therefore be potentially useful to rule out the presence of metabolically active disease in such patients, or serve as a second reader or decision support tool. FUNDING: National Institutes of Health-National Cancer Institute Cancer Center Support Grant.

Early identification and characterisation of stroke to support prehospital decision-making using artificial intelligence: a scoping review protocol.

INTRODUCTION: Stroke is a time-critical condition and one of the leading causes of mortality and disability worldwide. To decrease mortality and improve patient outcome by improving access to optimal treatment, there is an emerging need to improve the accuracy of the methods used to identify and characterise stroke in prehospital settings and emergency departments (EDs). This might be accomplished by developing computerised decision support systems (CDSSs) that are based on artificial intelligence (AI) and potential new data sources such as vital signs, biomarkers and image and video analysis. This scoping review aims to summarise literature on existing methods for early characterisation of stroke by using AI. METHODS AND ANALYSIS: The review will be performed with respect to the Arksey and O'Malley's model. Peer-reviewed articles about AI-based CDSSs for the characterisation of stroke or new potential data sources for stroke CDSSs, published between January 1995 and April 2023 and written in English, will be included. Studies reporting methods that depend on mobile CT scanning or with no focus on prehospital or ED care will be excluded. Screening will be done in two steps: title and abstract screening followed by full-text screening. Two reviewers will perform the screening process independently, and a third reviewer will be involved in case of disagreement. Final decision will be made based on majority vote. Results will be reported using a descriptive summary and thematic analysis. ETHICS AND DISSEMINATION: The methodology used in the protocol is based on information publicly available and does not need ethical approval. The results from the review will be submitted for publication in a peer-reviewed journal. The findings will be shared at relevant national and international conferences and meetings in the field of digital health and neurology.

Impact of ComBat Harmonization on PET Radiomics-Based Tissue Classification: A Dual-Center PET/MRI and PET/CT Study.

Our purpose was to determine whether ComBat harmonization improves 18F-FDG PET radiomics-based tissue classification in pooled PET/MRI and PET/CT datasets. Methods: Two hundred patients who had undergone 18F-FDG PET/MRI (2 scanners and vendors; 50 patients each) or PET/CT (2 scanners and vendors; 50 patients each) were retrospectively included. Gray-level histogram, gray-level cooccurrence matrix, gray-level run-length matrix, gray-level size-zone matrix, and neighborhood gray-tone difference matrix radiomic features were calculated for volumes of interest in the disease-free liver, spleen, and bone marrow. For individual feature classes and a multiclass radiomic signature, tissue was classified on ComBat-harmonized and unharmonized pooled data, using a multilayer perceptron neural network. Results: Median accuracies in training and validation datasets were 69.5% and 68.3% (harmonized), respectively, versus 59.5% and 58.9% (unharmonized), respectively, for gray-level histogram; 92.1% and 86.1% (harmonized), respectively, versus 53.6% and 50.0% (unharmonized), respectively, for gray-level cooccurrence matrix; 84.8% and 82.8% (harmonized), respectively, versus 62.4% and 58.3% (unharmonized), respectively, for gray-level run-length matrix; 87.6% and 85.6% (harmonized), respectively, versus 56.2% and 52.8% (unharmonized), respectively, for gray-level size-zone matrix; 79.5% and 77.2% (harmonized), respectively, versus 54.8% and 53.9% (unharmonized), respectively, for neighborhood gray-tone difference matrix; and 86.9% and 84.4% (harmonized), respectively, versus 62.9% and 58.3% (unharmonized), respectively, for radiomic signature. Conclusion: ComBat harmonization may be useful for multicenter 18F-FDG PET radiomics studies using pooled PET/MRI and PET/CT data.

Accelerating Prostate Diffusion-weighted MRI Using a Guided Denoising Convolutional Neural Network: Retrospective Feasibility Study.

PURPOSE: To investigate the feasibility of accelerating prostate diffusion-weighted imaging (DWI) by reducing the number of acquired averages and denoising the resulting image using a proposed guided denoising convolutional neural network (DnCNN). MATERIALS AND METHODS: Raw data from the prostate DWI scans were retrospectively gathered between July 2018 and July 2019 from six single-vendor MRI scanners. There were 103 datasets used for training (median age, 64 years; interquartile range [IQR], 11), 15 for validation (median age, 68 years; IQR, 12), and 37 for testing (median age, 64 years; IQR, 12). High b-value diffusion-weighted (hb DW) data were reconstructed into noisy images using two averages and reference images using all 16 averages. A conventional DnCNN was modified into a guided DnCNN, which uses the low b-value DW image as a guidance input. Quantitative and qualitative reader evaluations were performed on the denoised hb DW images. A cumulative link mixed regression model was used to compare the readers' scores. The agreement between the apparent diffusion coefficient (ADC) maps (denoised vs reference) was analyzed using Bland-Altman analysis. RESULTS: Compared with the original DnCNN, the guided DnCNN produced denoised hb DW images with higher peak signal-to-noise ratio (32.79 ± 3.64 [standard deviation] vs 33.74 ± 3.64), higher structural similarity index (0.92 ± 0.05 vs 0.93 ± 0.04), and lower normalized mean square error (3.9% ± 10 vs 1.6% ± 1.5) (P < .001 for all). Compared with the reference images, the denoised images received higher image quality scores from the readers (P < .0001). The ADC values based on the denoised hb DW images were in good agreement with the reference ADC values (mean ADC difference ranged from -0.04 to 0.02 × 10-3 mm2/sec). CONCLUSION: Accelerating prostate DWI by reducing the number of acquired averages and denoising the resulting image using the proposed guided DnCNN is technically feasible. Supplemental material is available for this article. © RSNA, 2020.

Introduction to Radiomics.

Radiomics is a rapidly evolving field of research concerned with the extraction of quantitative metrics-the so-called radiomic features-within medical images. Radiomic features capture tissue and lesion characteristics such as heterogeneity and shape and may, alone or in combination with demographic, histologic, genomic, or proteomic data, be used for clinical problem solving. The goal of this continuing education article is to provide an introduction to the field, covering the basic radiomics workflow: feature calculation and selection, dimensionality reduction, and data processing. Potential clinical applications in nuclear medicine that include PET radiomics-based prediction of treatment response and survival will be discussed. Current limitations of radiomics, such as sensitivity to acquisition parameter variations, and common pitfalls will also be covered.

DeepPET: A deep encoder-decoder network for directly solving the PET image reconstruction inverse problem.

The purpose of this research was to implement a deep learning network to overcome two of the major bottlenecks in improved image reconstruction for clinical positron emission tomography (PET). These are the lack of an automated means for the optimization of advanced image reconstruction algorithms, and the computational expense associated with these state-of-the art methods. We thus present a novel end-to-end PET image reconstruction technique, called DeepPET, based on a deep convolutional encoder-decoder network, which takes PET sinogram data as input and directly and quickly outputs high quality, quantitative PET images. Using simulated data derived from a whole-body digital phantom, we randomly sampled the configurable parameters to generate realistic images, which were each augmented to a total of more than 291,000 reference images. Realistic PET acquisitions of these images were simulated, resulting in noisy sinogram data, used for training, validation, and testing the DeepPET network. We demonstrated that DeepPET generates higher quality images compared to conventional techniques, in terms of relative root mean squared error (11%/53% lower than ordered subset expectation maximization (OSEM)/filtered back-projection (FBP), structural similarity index (1%/11% higher than OSEM/FBP), and peak signal-to-noise ratio (1.1/3.8 dB higher than OSEM/FBP). In addition, we show that DeepPET reconstructs images 108 and 3 times faster than OSEM and FBP, respectively. Finally, DeepPET was successfully applied to real clinical data. This study shows that an end-to-end encoder-decoder network can produce high quality PET images at a fraction of the time compared to conventional methods.

Initial performance studies of a wearable brain positron emission tomography camera based on autonomous thin-film digital Geiger avalanche photodiode arrays.

Using analytical and Monte Carlo modeling, we explored performance of a lightweight wearable helmet-shaped brain positron emission tomography (PET), or BET camera, based on thin-film digital Geiger avalanche photodiode arrays with Lutetium-yttrium oxyorthosilicate (LYSO) or [Formula: see text] scintillators for imaging in vivo human brain function of freely moving and acting subjects. We investigated a spherical cap BET and cylindrical brain PET (CYL) geometries with 250-mm diameter. We also considered a clinical whole-body (WB) LYSO PET/CT scanner. The simulated energy resolutions were 10.8% (LYSO) and 3.3% ([Formula: see text]), and the coincidence window was set at 2 ns. The brain was simulated as a water sphere of uniform F-18 activity with a radius of 100 mm. We found that BET achieved [Formula: see text] better noise equivalent count (NEC) performance relative to the CYL and [Formula: see text] than WB. For 10-mm-thick [Formula: see text] equivalent mass systems, LYSO (7-mm thick) had [Formula: see text] higher NEC than [Formula: see text]. We found that [Formula: see text] scintillator crystals achieved [Formula: see text] full-width-half-maximum spatial resolution without parallax errors. Additionally, our simulations showed that LYSO generally outperformed [Formula: see text] for NEC unless the timing resolution for [Formula: see text] was considerably smaller than that presently used for LYSO, i.e., well below 300 ps.

Dynamic PET simulator via tomographic emission projection for kinetic modeling and parametric image studies.

PURPOSE: To develop and evaluate a fast and simple tool called dpetstep (Dynamic PET Simulator of Tracers via Emission Projection), for dynamic PET simulations as an alternative to Monte Carlo (MC), useful for educational purposes and evaluation of the effects of the clinical environment, postprocessing choices, etc., on dynamic and parametric images. METHODS: The tool was developed in matlab using both new and previously reported modules of petstep (PET Simulator of Tracers via Emission Projection). Time activity curves are generated for each voxel of the input parametric image, whereby effects of imaging system blurring, counting noise, scatters, randoms, and attenuation are simulated for each frame. Each frame is then reconstructed into images according to the user specified method, settings, and corrections. Reconstructed images were compared to MC data, and simple Gaussian noised time activity curves (GAUSS). RESULTS: dpetstep was 8000 times faster than MC. Dynamic images from dpetstep had a root mean square error that was within 4% on average of that of MC images, whereas the GAUSS images were within 11%. The average bias in dpetstep and MC images was the same, while GAUSS differed by 3% points. Noise profiles in dpetstep images conformed well to MC images, confirmed visually by scatter plot histograms, and statistically by tumor region of interest histogram comparisons that showed no significant differences (p < 0.01). Compared to GAUSS, dpetstep images and noise properties agreed better with MC. CONCLUSIONS: The authors have developed a fast and easy one-stop solution for simulations of dynamic PET and parametric images, and demonstrated that it generates both images and subsequent parametric images with very similar noise properties to those of MC images, in a fraction of the time. They believe dpetstep to be very useful for generating fast, simple, and realistic results, however since it uses simple scatter and random models it may not be suitable for studies investigating these phenomena. dpetstep can be downloaded free of cost from https://github.com/CRossSchmidtlein/dPETSTEP.

PETSTEP: Generation of synthetic PET lesions for fast evaluation of segmentation methods.

PURPOSE: This work describes PETSTEP (PET Simulator of Tracers via Emission Projection): a faster and more accessible alternative to Monte Carlo (MC) simulation generating realistic PET images, for studies assessing image features and segmentation techniques. METHODS: PETSTEP was implemented within Matlab as open source software. It allows generating three-dimensional PET images from PET/CT data or synthetic CT and PET maps, with user-drawn lesions and user-set acquisition and reconstruction parameters. PETSTEP was used to reproduce images of the NEMA body phantom acquired on a GE Discovery 690 PET/CT scanner, and simulated with MC for the GE Discovery LS scanner, and to generate realistic Head and Neck scans. Finally the sensitivity (S) and Positive Predictive Value (PPV) of three automatic segmentation methods were compared when applied to the scanner-acquired and PETSTEP-simulated NEMA images. RESULTS: PETSTEP produced 3D phantom and clinical images within 4 and 6 min respectively on a single core 2.7 GHz computer. PETSTEP images of the NEMA phantom had mean intensities within 2% of the scanner-acquired image for both background and largest insert, and 16% larger background Full Width at Half Maximum. Similar results were obtained when comparing PETSTEP images to MC simulated data. The S and PPV obtained with simulated phantom images were statistically significantly lower than for the original images, but led to the same conclusions with respect to the evaluated segmentation methods. CONCLUSIONS: PETSTEP allows fast simulation of synthetic images reproducing scanner-acquired PET data and shows great promise for the evaluation of PET segmentation methods.

A Monte Carlo study of the dependence of early frame sampling on uncertainty and bias in pharmacokinetic parameters from dynamic PET.

UNLABELLED: Compartmental modeling of dynamic PET data enables quantification of tracer kinetics in vivo, through the calculated model parameters. In this study, we aimed to investigate the effect of early frame sampling and reconstruction method on pharmacokinetic parameters obtained from a 2-tissue model, in terms of bias and uncertainty (SD). METHODS: The GATE Monte Carlo software was used to simulate 2 × 15 dynamic 3'-deoxy-3'-(18)F-fluorothymidine ((18)F-FLT) brain PET studies, typical in terms of noise level and kinetic parameters. The data were reconstructed by both 3-dimensional (3D) filtered backprojection with reprojection (3DRP) and 3D ordered-subset expectation maximization (OSEM) into 6 dynamic image sets with different early frame durations of 1, 2, 4, 6, 10, and 15 s. Bias and SD were evaluated for fitted parameter estimates, calculated from regions of interest. RESULTS: The 2-tissue-model parameter estimates K1, k2, and fraction of arterial blood in tissue depended on early frame sampling, and a sampling of 6-15 s generally minimized bias and SD. The shortest sampling of 1 s yielded a 25% and 42% larger bias than the other schemes, for 3DRP and OSEM, respectively, and a parameter uncertainty that was 10%-70% higher. The schemes from 4 to 15 s were generally not significantly different in regards to bias and SD. Typically, the reconstruction method 3DRP yielded less frame-sampling dependence and less uncertain results, compared with OSEM, but was on average more biased. CONCLUSION: Of the 6 sampling schemes investigated in this study, an early frame duration of 6-15 s generally kept both bias and uncertainty to a minimum, for both 3DRP and OSEM reconstructions. Very-short frames of 1 s should be avoided because they typically resulted in the largest parameter bias and uncertainty. Furthermore, 3DRP may be preferred over OSEM for short frames with poor statistics.

Compartment modeling of dynamic brain PET--the impact of scatter corrections on parameter errors.

PURPOSE: The aim of this study was to investigate the effect of scatter and its correction on kinetic parameters in dynamic brain positron emission tomography (PET) tumor imaging. The 2-tissue compartment model was used, and two different reconstruction methods and two scatter correction (SC) schemes were investigated. METHODS: The gate Monte Carlo (MC) software was used to perform 2 × 15 full PET scan simulations of a voxelized head phantom with inserted tumor regions. The two sets of kinetic parameters of all tissues were chosen to represent the 2-tissue compartment model for the tracer 3'-deoxy-3'-((18)F)fluorothymidine (FLT), and were denoted FLT1 and FLT2. PET data were reconstructed with both 3D filtered back-projection with reprojection (3DRP) and 3D ordered-subset expectation maximization (OSEM). Images including true coincidences with attenuation correction (AC) and true+scattered coincidences with AC and with and without one of two applied SC schemes were reconstructed. Kinetic parameters were estimated by weighted nonlinear least squares fitting of image derived time-activity curves. Calculated parameters were compared to the true input to the MC simulations. RESULTS: The relative parameter biases for scatter-eliminated data were 15%, 16%, 4%, 30%, 9%, and 7% (FLT1) and 13%, 6%, 1%, 46%, 12%, and 8% (FLT2) for K1, k2, k3, k4, Va, and Ki, respectively. As expected, SC was essential for most parameters since omitting it increased biases by 10 percentage points on average. SC was not found necessary for the estimation of Ki and k3, however. There was no significant difference in parameter biases between the two investigated SC schemes or from parameter biases from scatter-eliminated PET data. Furthermore, neither 3DRP nor OSEM yielded the smallest parameter biases consistently although there was a slight favor for 3DRP which produced less biased k3 and Ki estimates while OSEM resulted in a less biased Va. The uncertainty in OSEM parameters was about 26% (FLT1) and 12% (FLT2) larger than for 3DRP although identical postfilters were applied. CONCLUSIONS: SC was important for good parameter estimations. Both investigated SC schemes performed equally well on average and properly corrected for the scattered radiation, without introducing further bias. Furthermore, 3DRP was slightly favorable over OSEM in terms of kinetic parameter biases and SDs.

Monolithic Bragg-locked Nd:GdVO4 laser.

We present a monolithic single-longitudinal-mode laser based on Nd:GdVO(4) and a volume Bragg grating. The laser at 1066 nm had a bandwidth below 40 MHz at a power of 30 mW. With temperature, the laser frequency could be continuously tuned without mode hops over a range of 80 GHz. The demonstrated laser design is very compact and simple and can be used to lock the laser wavelength anywhere in the gain spectrum.