Portal de Periódicos da CAPES

The Dawn of a New Era in Low-Dose PET Imaging

2018; Radiological Society of North America; Volume: 290; Issue: 3 Linguagem: Inglês

10.1148/radiol.2018182573

1527-1315

Ciprian Catana,

Radiomics and Machine Learning in Medical Imaging

HomeRadiologyVol. 290, No. 3 PreviousNext Reviews and CommentaryFree AccessEditorialThe Dawn of a New Era in Low-Dose PET ImagingCiprian Catana Ciprian Catana Author AffiliationsFrom the Department of Radiology, Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, A.A. Martinos Center, 149 13th St, Room 2.301, Charlestown, MA 02129.Address correspondence to the author (e-mail: [email protected]).Ciprian Catana Published Online:Dec 11 2018https://doi.org/10.1148/radiol.2018182573MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Chen et al in this issue.IntroductionOne of the most important tradeoffs to be considered in PET imaging is the one between image quality and radiation exposure. Improving image quality is desirable for virtually all clinical and research applications, while minimizing radiation exposure is needed to reduce the risks associated with ionizing radiation. Addressing these two apparently contradictory requirements has been a continued balancing act in the field, especially given the ever-increasing concerns related to radiation exposure in patients requiring multiple examinations (eg, for treatment monitoring in oncology) or those at a higher lifetime risk for developing cancer (eg, pediatric patients). Furthermore, reducing radiation exposure in PET studies without sacrificing image quality could potentially expand its utility beyond current applications.For a photon-count limited modality such as PET, noise degrades the image quality and is always present in the data. The PET scanner detects only a small fraction of the emitted photons because its sensitivity is a few percent at best. Acquisition time must be kept short for patient comfort. Therefore, improving the signal-to-noise ratio requires administration of a larger amount of radiotracer, which exposes the patient to a higher amount of radiation exposure. Minimizing this exposure without detrimentally affecting the image quality can be achieved by improving the performance of the hardware and/or using software techniques.For a long time, equipment manufacturers have focused mainly on improving the sensitivity of PET scanners in order to detect a larger fraction of the annihilation photons. This can be accomplished by extending the axial field of view and reducing the scanner diameter (to maximize the geometric coverage), using thick detectors with high linear attenuation coefficient (to improve the photon detection efficiency), acquiring the data in three-dimensional mode, and using higher performance detectors to exclude unwanted events and spatially constrain the positron annihilation location. The scanner recently developed by the EXPLORER consortium (https://explorer.ucdavis.edu/) is the perfect example of using a hardware approach to push the limits of sensitivity, allowing a 40-fold increase in effective sensitivity for entire-body imaging (compared with state-of-the-art whole-body PET scanners) and a fivefold increase when imaging the brain (1). Cherry et al even discussed the possibility of increasing the sensitivity 200-fold by improving the detector efficiency and timing resolution.In parallel, numerous signal-processing techniques have been developed to minimize the contribution of noise (and other limitations) in order to improve PET image quality. These techniques include standard smoothing routinely applied after image reconstruction, image denoising techniques, and approaches that use information obtained from another modality. The development of integrated PET and CT or MRI scanners was a logical next step as it allowed the combined use of hardware and software developments. To give just one example from the previous work of Chen et al, an efficient MR-assisted PET data optimization approach was recently proposed, allowing spatially and temporally correlated MR data acquired simultaneously with PET data to be used for attenuation correction, motion compensation, and anatomy-aided PET image reconstruction (2). Integrating PET with MRI instead of CT also inherently reduces radiation exposure, an additional benefit of this approach.More recently, machine learning approaches have been used successfully for a range of applications in health care. When applied to medical imaging, these methods enable us to build upon the above-mentioned technological and methodological advances and further leverage the multimodality data. In this issue of Radiology, Chen et al (3) propose a method to reduce the radiotracer dose by using the simultaneously acquired MRI morphologic information and a machine learning approach to synthesize diagnostic quality from low-dose PET images. Prior frameworks for generating standard-dose from low-dose PET images and complementary anatomic MRI were based on multiple approaches. These include regression forests (4), multilevel cananonical correlation analysis (5), mapping-based sparsed representation (6), semisupervised tripled dictionary learning (7), as well as deep learning approaches using deep auto-context convolutional neural networks (8), and three-dimensional generative adversarial networks (9). In addition to having names that might be unfamiliar to clinician readers, these machine learning frameworks were all tested in a similar scenario—to predict standard-dose fluorine 18 (18F) fluorodeoxyglucose (FDG) brain PET images from fourfold-lower-dose PET images and anatomic MRI data collected from individuals with normal brain anatomy and glucose metabolism.Instead of this similar scenario, Chen et al proposed a more dramatic improvement, reducing the dose 100-fold (3). While the same group previously reported on the potential to reduce the 18F-FDG dose 200-fold in glioblastoma patients (10), the current study is the first (to the author's knowledge) to demonstrate that such a method can be applied to a different radiotracer, 18F-florbetaben, one that is arguably more relevant than 18F-FDG for PET studies at earlier stages in the progression of Alzheimer disease. Furthermore, in addition to normal controls, Chen et al also included patients with mild cognitive impairment and patients with neurodegenerative disease. A total of 40 multimodal data sets were retrospectively analyzed. Convolutional neural networks were trained using the low-dose PET images (generated by randomly extracting 1% of the events from the full-dose raw-emission data) and the multicontrast MR images (consisting of T1-, T2-, and T2 fluid-attenuated inversion recovery–weighted images) to generate PET-only and PET-plus-MRI models. The synthetic images obtained with these models were compared qualitatively and quantitatively, and clinical readings were performed to assess the amyloid uptake status using the full-dose images as the ground truth. The image quality was substantially improved compared with the low-dose images, particularly when using the PET-plus-MRI model. This model also had high accuracy in assessing the amyloid status and demonstrated low bias and small variance in the estimation of standard uptake value ratios compared with the full-dose images. Overall, the results demonstrate the feasibility of the proposed deep learning approach in the context of amyloid imaging, while the possibility of reducing the dose suggests PET could be used for applications such as screening of younger patients at risk for developing Alzheimer disease (3). This would be extremely important, as most clinical trials fail in later stages and the focus of therapeutic developments has been slowly shifting to earlier stages of disease.Although the sample size of 40 data sets was relatively limited, it included sufficient patient subpopulations to suggest that the method could be used for 18F-florbetaben amyloid imaging in general and that a similar approach could be used to train convolutional neural networks for synthesizing images for other radiotracers. However, additional studies using appropriate training data will be needed to demonstrate these points. In fact, it would be interesting to test whether these methods could be used in a disease-agnostic manner (although they will likely be radiotracer-specific). Another limitation of this study was that the contribution of the various MR contrasts was not individually assessed. If a model trained using the low-dose PET and only the T1-weighted MR images provided similar results, the acquisition time could be further reduced to minimize the potential negative effects of head motion and improve patient compliance. Finally, the 100-fold dose reduction chosen by the authors was arbitrary and the true limits for specific applications should be assessed in future studies.Chen et al compared the ultra–low-dose radiation exposure to that from a transcontinental flight (3), which would make PET an imaging modality with low radiation exposure. While their results are remarkable and could have immediate clinical implications, this might be just the beginning of a new era in PET imaging. Indeed, in the not-too-distant future, the prospect of combining both the hardware and software approaches (ie, ultra–high-sensitivity PET scanners, simultaneously acquired MR data, and machine learning) to improve the PET image quality could become a reality. This combination would allow further dramatic reductions of upwards of 10 000-fold in the dose required for obtaining excellent-quality images. To put this into perspective, this level would be comparable to the yearly exposure from natural sources such as the water (with radium 226) we drink, the bananas (with potassium 40) we eat, the natural gas (with radon 222) we use to heat our homes, or all the other radionuclides that are already part of our bodies. In fact, one would be twice more likely to win the lottery than to develop cancer throughout his or her lifetime due to the radiation exposure from a single PET examination. These advances may ultimately transform PET into a virtually nonionizing-radiation imaging modality and would give us the opportunity to truly unleash its potential.Disclosures of Conflicts of Interest: C.C. disclosed no relevant relationships.References1. Cherry SR, Jones T, Karp JS, Qi J, Moses WW, Badawi RD. Total-body PET: maximizing sensitivity to create new opportunities for clinical research and patient care. J Nucl Med 2018;59(1):3–12. Crossref, Medline, Google Scholar2. Chen KT, Salcedo S, Gong K, et al. An efficient approach to perform MR-assisted PET data optimization in simultaneous PET/MR neuroimaging studies. J Nucl Med 2018 Jun 22 [Epub ahead of print]. Google Scholar3. Chen KT, Gong E, Macruz F. Ultra-low-dose 18 F-florbetaben amyloid PET imaging using deep learning with multi-contrast MRI inputs. Radiology 2019;290:649–656. Link, Google Scholar4. Kang J, Gao Y, Shi F, Lalush DS, Lin W, Shen D. P online rediction of standard-dose brain PET image by using MRI and low-dose brain [18F]FDG PET images. Med Phys 2015;42(9):5301–5309. Crossref, Medline, Google Scholar5. Le An, Pei Zhang, Adeli E, et al. Multi-level canonical correlation analysis for standard-dose PET image estimation. IEEE Trans Image Process 2016;25(7):3303–3315. Crossref, Medline, Google Scholar6. Wang Y, Zhang P, An L, et al. Predicting standard-dose PET image from low-dose PET and multimodal MR images using mapping-based sparse representation. Phys Med Biol 2016;61(2):791–812. Crossref, Medline, Google Scholar7. Wang Y, Ma G, An L, et al. Semisupervised tripled dictionary learning for standard-dose PET image prediction using low-dose PET and multimodal MRI. IEEE Trans Biomed Eng 2017;64(3):569–579. Crossref, Medline, Google Scholar8. Xiang L, Qiao Y, Nie D, An L, Wang Q, Shen D. Deep auto-context convolutional neural networks for standard-dose PET image estimation from low-dose PET/MRI. Neurocomputing 2017;267:406–416. Crossref, Medline, Google Scholar9. Wang Y, Yu B, Wang L, et al. 3D conditional generative adversarial networks for high-quality PET image estimation at low dose. Neuroimage 2018;174:550–562. Crossref, Medline, Google Scholar10. Xu J, Gong E, Pauly J, Zaharchuk G. 200x Low-dose PET Reconstruction using Deep Learning. http://arxiv.org/abs/1712.04119. Published 2017. Accessed November 8, 2018. Google ScholarArticle HistoryReceived: Nov 8 2018Revision requested: Nov 12 2018Revision received: Nov 12 2018Accepted: Nov 14 2018Published online: Dec 11 2018Published in print: Mar 2019 FiguresReferencesRelatedDetailsCited ByFunctional NeuroradiologyCiprianCatana, ChristinSander, A. GregorySorensen, Bruce R.Rosen2023Nuclear Medicine and Molecular ImagingJohn O.Prior, GillesAllenbach2022Evaluation of a High-Sensitivity Organ-Targeted PET CameraJustinStiles, BrandonBaldassi, OleksandrBubon, HarutyunPoladyan, VivianneFreitas, AnabelScaranelo, Anna MarieMulligan, MichaelWaterston, AllaReznik2022 | Sensors, Vol. 22, No. 13Progress and perspectives in total body PET systems instrumentationStefaanVandenberghe2022 | Bio-Algorithms and Med-Systems, Vol. 17, No. 4Positron emission tomography in multiple sclerosis — straight to the targetBenedettaBodini, MatteoTonietto, LauraAiras, BrunoStankoff2021 | Nature Reviews Neurology, Vol. 17, No. 11Task-evoked simultaneous FDG-PET and fMRI data for measurement of neural metabolism in the human visual cortexSharna D.Jamadar, ShenjunZhong, AlexandraCarey, RichardMcIntyre, Phillip G. D.Ward, AlexFornito, MalinPremaratne, NJon Shah, KieranO'Brien, DanielStäb, ZhaolinChen, Gary F.Egan2021 | Scientific Data, Vol. 8, No. 1Improved low-dose positron emission tomography image reconstruction using deep learned priorXinhuiWang, LongZhou, YaofaWang, HaochuanJiang, HongweiYe2021 | Physics in Medicine & Biology, Vol. 66, No. 11PET/MRI in paediatric diseaseTaraSepehrizadeh, IanJong, MichaelDeVeer, AtulMalhotra2021 | European Journal of Radiology, Vol. 144The Evolution of Image Reconstruction in PETKuangGong, KyungsangKim, JiananCui, DufanWu, QuanzhengLi2021 | PET Clinics, Vol. 16, No. 4Künstliche Intelligenz in der HybridbildgebungChristianStrack, RobertSeifert, JensKleesiek2020 | Der Radiologe, Vol. 60, No. 5Machine learning in quantitative PET: A review of attenuation correction and low-count image reconstruction methodsTongheWang, YangLei, YaboFu, Walter J.Curran, TianLiu, Jonathon A.Nye, XiaofengYang2020 | Physica Medica, Vol. 76State of the art in total body PETStefaanVandenberghe, PawelMoskal, Joel S.Karp2020 | EJNMMI Physics, Vol. 7, No. 1Reduction of the fluorine-18-labeled fluorodeoxyglucose dose for clinically dedicated breast positron emission tomographyYokoSatoh, TetsuroSekine, YoshieOmiya, HiroshiOnishi, UtarohMotosugi2019 | EJNMMI Physics, Vol. 6, No. 1Development of Dedicated Brain PET Imaging Devices: Recent Advances and Future PerspectivesCiprianCatana2019 | Journal of Nuclear Medicine, Vol. 60, No. 8Accompanying This ArticleUltra–Low-Dose 18F-Florbetaben Amyloid PET Imaging Using Deep Learning with Multi-Contrast MRI InputsDec 11 2018RadiologyRecommended Articles Ultra–Low-Dose 18F-Florbetaben Amyloid PET Imaging Using Deep Learning with Multi-Contrast MRI InputsRadiology2018Volume: 290Issue: 3pp. 649-656Generative Adversarial Networks: A Primer for RadiologistsRadioGraphics2021Volume: 41Issue: 3pp. 840-857MRI Techniques to Decrease Imaging Times in ChildrenRadioGraphics2020Volume: 40Issue: 2pp. 485-502Robust Motion Correction Strategy for Structural MRI in Unsedated Children Demonstrated with Three-dimensional Radial MPnRAGERadiology2018Volume: 289Issue: 2pp. 509-516Overview of Noninterpretive Artificial Intelligence Models for Safety, Quality, Workflow, and Education Applications in Radiology PracticeRadiology: Artificial Intelligence2022Volume: 4Issue: 2See More RSNA Education Exhibits IAEA Atlas on Quality Control and Image Artefacts in SPECT/CTDigital Posters2019A Rational Approach to Pediatric MRI OptimizationDigital Posters2020Synthetic MRI, STAGE, and MR Acceleration through Deep-Learning: A Comparison of Time-Efficient MR Techniques in NeuroimagingDigital Posters2022 RSNA Case Collection Creutzfeldt-Jakob DiseaseRSNA Case Collection2021Semantic dementiaRSNA Case Collection2020Non-accidental anoxic brain injury RSNA Case Collection2021 Vol. 290, No. 3 Metrics Altmetric Score PDF download