Professional Documents
Culture Documents
Nuclear Medicine
Gaspar Delso, PhD,* Davide Cirillo, PhD,† Joshua D Kaggie, PhD,z
Alfonso Valencia, PhD Prof.,† Ur Metser, MD Prof.,x and Patrick Veit-Haibach, MD Prof.x
Artificial intelligence (AI) is an overarching term for a multitude of technologies which are
currently being discussed and introduced in several areas of medicine and in medical imag-
ing specifically. There is, however, limited literature and information about how AI techni-
ques can be integrated into the design of clinical imaging trials. This article will present
several aspects of AI being used in trials today and how imaging departments and espe-
cially nuclear medicine departments can prepare themselves to be at the forefront of AI-
driven clinical trials. Beginning with some basic explanation on AI techniques currently
being used and existing challenges of its implementation, it will also cover the logistical
prerequisites which have to be in place in nuclear medicine departments to participate suc-
cessfully in AI-driven clinical trials.
Semin Nucl Med 51:112-119 © 2020 Elsevier Inc. All rights reserved.
112 https://doi.org/10.1053/j.semnuclmed.2020.09.003
0001-2998/© 2020 Elsevier Inc. All rights reserved.
How to Design AI-Driven Clinical Trials in Nuclear Medicine 113
identification of potential investigators, including spe- companies and several departments of the Massachusetts
cialists, committees and scientific boards, as well as the Institute of Technology to automate small molecule discov-
selection of high-performance sites, based on budget, ery and synthesis. Similar consortia are being created, albeit
qualifications and patient population, are planning at a slower pace, for the field of radiopharmaceutical devel-
aspects of clinical trials that can be informed and opti- opment (eg, Isotope4life).
mized by AI. Moreover, AI can reduce the impact of Another well-known limitation of modern AI is its depen-
human error in data collection, discover trends and pin- dency on vast amounts of training data. Clinical trial data is
point relevant insights, and source quality information especially sensitive as it is connected to personal health infor-
to optimize production, supply chain and logistics. For mation that is heavily protected against misuse. In practical
instance, pharmaceutical manufacturers consider AI terms, hospitals differ in their willingness to integrate new AI
essential in specific areas such as cost-efficiently optimi- systems as these can cause risk if a patient’s information is
zation of compound administration and dosing leaked or misdiagnosed by these systems. Hospitals can also
regimens.29 differ in their business models, being for-profit or non-for-
profit, which can cause varied engagement with commercial
institutions. Data acquisition is one of the most significant
As for nuclear medicine, several areas have benefited from AI hurdles in AI training. At present, AI development may have
applications, including nuclear cardiology, oncology, and neu- intellectual property surrounding their code, their data, or
rology.30-37 As with other imaging modalities, the most visible future developments, that a hospital or commercial entity
aspect has been a proliferation of segmentation and classifica- may want to protect. Especially in imaging and nuclear medi-
tion applications for diagnostic support.38 However, AI is also cine, there are large variations across institutions even in the
being applied in creative new ways: image quality enhancement same jurisdiction concerning acquisition of imaging data,
from low dose acquisitions39; reconstruction40,41; attenuation reconstruction, transfer and storage.48-54 Even in optimally
correction for hybrid imaging42; spatial normalization43; multi- functioning consortia adhering to strict data harmonization
modal data analysis44; predicting patient response to therapy45; policies, intrinsic differences in data quality (eg, due to scan-
implementation of AI-powered imaging biomarkers,46 opening ner performance) may impair the use of common AI models
up the possibility of large-scale population screening with a to extract and analyze the resulting biomarkers.55 Addition-
radical impact in clinical trial improvement. ally, nuclear medicine data as semiquantitative (and partly
qualitative-only) modality, is prone to subjective interpreta-
tion and thus, automated mining of data currently poses a
Challenges of integrating AI into comparability and curation problem, introducing a potential
bias for clinical trials.
trials Significant efforts, driven by both public and private
Despite the undeniable potential of AI methods to streamline investment, have been dedicated in recent years to the
clinical trials, a word of caution is required. The recent prolif- harmonization and standardization of biomedical data, the
eration of enthusiastic reports about the topic, often listing networking of databases and the development of open-
inspiring but episodic success stories, may offer a misleading source data management tools. Arguably, the biomedical
picture of the considerable amount of work still required to part of this effort56-59 is far more advanced than its healthcare
enable this vision. In the following sections we will focus on counterpart, with FAIR60 and AAI61 standards in place and
the case of Nuclear Medicine departments participating in implemented and journals, agencies, and institutions com-
clinical trials driven by pharmaceutical and medical technol- mitted to make data openly accessible. While their success
ogy companies. gives us reason to hope for a similar outcome in the medical
While the extraordinary leap forward of the AI field over field, there is still much work ahead.
the past ten years has been of a technical nature, it is impor- Natural language processing algorithms have only recently
tant to realize that the requirements for its successful integra- started to curate into structured, searchable databases the
tion in practice are unquestionably multidisciplinary. Drug staggering amount of usable data encoded in physician’s
and medical device manufacturers need to establish partner- notes, which may still not be digital or easily interpreted by
ships with academical institutions, clinical centers and ser- external parties. Challenges are inherent with data collection
vice providers in order to secure the required expertise and as new discovery techniques or diseases can emerge that may
infrastructure to support these new technologies. This com- not harmonize with data from older trials. Indeed, proven
prises, from an in-depth understanding of the principles of clinical care often lags behind new methods as the burden of
AI in order to select the best suited model for a given job, to proof is high for both clinical methods and drug trials in
the ability to recruit and process patients for the trial, to the order to reduce harms to the patients. With these constantly
technical means to provide long-term storage and curation of developing new methods, companies may also have con-
the resulting data. Good examples of efforts toward this goal stantly developing proprietary and changing formats that AI
are the MELLODDY Consortium,47 using part of the chemi- models are ill-equipped to handle.
cal libraries of 10 pharmaceutical firms to create a deep learn- The technical, administrative, and legal resources required
ing-based drug discovery service; or the MLPDS consortium, to expand AI efforts across clinical centers and even country
a collaboration between biotechnology and pharma borders, addressing such issues as harmonization, data
116 G. Delso et al.
ownership and protection from malicious access, can become Research Ethics evaluation process, which also captures
daunting, although not impossible.62,63 This is discussed in Nuclear Medicine specific topics like radiation burden and
“Clinical and ethical considerations of AI-enhanced trials.” radiopharmaceutical safety.
Imaging in nuclear medicine is particularly challenging
from the point of view of AI: costly examinations with rela-
Departmental prerequisites for tively low throughput and involving ionizing radiation. The
field of synthetic data generation offers an opportunity in
AI implementation this case.66 A key to best take advantage of synthetic data
Pharmaceutical and medical device companies are actively generation for AI purposes is controlling the image genera-
looking for partnerships that will allow them to enhance their tion model. Indeed, imaging modalities in nuclear medicine
clinical trials with promising new technologies. There are often involve a complex reconstruction process to transform
several steps that clinical sites can undertake to enhance their the data acquired by the scanner into human-readable
eligibility as participants of those trials. These are, for the images. Access to this reconstruction and its internal parame-
most part, as self-evident as going paperless and, in many ters, as well as a firm grasp of the underlying physics, will
cases, can take just as much effort and resources. We provide provide a huge boost to the amount and quality of the syn-
in this section a summary of those steps, with a focus on thetic data that can be produced.
Nuclear Medicine departments. There are many good reasons for the adoption of an elec-
There is a technical aspect to this transformation, to ensure tronic health record (EHR) software, too many to enumerate
that the more general hardware and software infrastructure in this article or to advise on a specific implementation
requirements are satisfied. But, more importantly, there is a (locally hosted vs cloud-based, proprietary vs open, etc.)
logistic and cultural aspect in the drive towards the generation Two important considerations are worth mentioning, when
of curated, accessible, standardized and, in summary, AI- considering such a move: first, this is an effort that will
friendly data. It might be objected that AI algorithms have require the engagement of the department staff, preferably
been demonstrated to successfully extract meaningful informa- with someone to champion it. Second, make sure to be aware
tion from unstructured datasets (a commercial example could of the data model used by the EHRs. There are several pro-
be the web-based healthcare data mining tool by Roam Analyt- posed standards to structure health record data and stan-
ics, San Mateo, CA). However feasible, unstructured data man- dardize database querying and export, including Common
agement remains an inefficient alternative to proper data Data Models intended to enable the combination of data
collection and curation practices. In any case, as a general from multiple sources (eg, Sentinel, OMOP, PCORNet, Vir-
rule, the better the data are organized and structured, the eas- tual Data Warehouse). While we do not wish to advocate for
ier future applications will be, AI-based and otherwise. one specifically, it is likely that clinical trial organizers will
An advisable first step would be cataloguing the various prioritize sites with readily accessible data stored following a
data sources controlled by the department and implementing well-known data model. Initiatives such as the Observational
protocols to manage their lifecycle.64,65 This includes all data Health Data Sciences and Informatics (OHDSI) and the Euro-
typically included in medical records (personal statistics, pean Health Data & Evidence Network (EHDEN) are good
demographics, medical history, medication and allergies, sources of information on this respect.
vital signs, test results, imaging, billing information, etc.), Several recommendations are valid before engaging in any
but also data about the facilities (medical devices, isotope, kind of clinical trial: the adoption of a document manage-
and tracer production), personnel and expertise, research ment system and the associated practices are highly advis-
protocols, ethical approval documentation, informed consent able; when data measurements are generated, calibration and
forms, information, and divulgation materials, etc. quality control measurements will have to be periodically
For Nuclear Medicine departments specifically, this means recorded; considering the expensive nature of the data, it is
that there ideally needs to exist a centralized, standardized important to also consider the potential effect of unforeseen
system which can capture the multiple processes which are disastrous events, like a fire, such that data or models should
unique for Nuclear Medicine and mine them for predictive/ be backed up in multiple locations. This work is typically
diagnostic/prescriptive analytics. Data sources to be captured done by clinical research coordinators or physicians them-
in here are, for example, the radiopharmacy. The steps to be selves (depending on the jurisdiction), trained profoundly in
documented in a standardized way range from the produc- GCP trial coordination and documentation. For nuclear
tion processes (several steps for low-volume tracers are not medicine studies specifically, the correct and compliant
automated), the status and delivery of precursors, the equip- documentation of the most sensitive data (ie, radiopharma-
ment used (ie, reagents and hotcells), to the quality control ceutical or contrast media dosages, acquisition time, recon-
and shipping logistics. For the imaging equipment, a separa- struction parameters) are of utmost importance. Here, a
tion of records might be needed that is, for preclinical vs. standardized documentation system (paper based earlier,
clinical systems. Also, differences between research acquisi- currently moving into electronic system administration) is
tion vs standard clinical acquisition (injected tracer doses, required by the authorities and is vigilantly monitored. AI
acquisition time, reconstruction) must be monitored and techniques and machine learning algorithm are prone to
documented to be comprehensible and minable in an effi- deliver significantly improved results when electronic based
cient way. The documentation can be even extended to the and standardized systems are available as a database.
How to Design AI-Driven Clinical Trials in Nuclear Medicine 117
Data consistency is a key factor affecting the performance following topics and engage their ethical and legal depart-
of trained methods. Sites will have to be particularly vigilant ments in the discussion:
to ensure that no significant changes are introduced in the
data generation pathway during a study (eg, from the way - Security: Gathering large amounts of well-structured
the patient weight is measured to the activity of patients dur- data in electronic repositories carries an inevitable risk
ing radiotracer uptake time). This may include events such of exposing confidential patient information, either
as equipment servicing or upgrades, re-calibration, changes accidentally or through a malicious security breach.
in compound production or supplier, new personnel, etc. This is particularly critical in the case of Nuclear Medi-
This is particularly critical when AI models of the “black cine, given the severity of the conditions involved. Per-
box” type are used, as they are particularly sensitive to data sonnel training and strict adherence to the relevant
inconsistencies and biases during their training. data protection regulations (GDPR, CCPA, etc.) will go
Personnel awareness and involvement will be a key factor a long way towards minimizing these risks. In general,
in order to ensure not just data consistency but, more impor- straightforward, well-designed data pipelines and pro-
tantly, provable consistency, in the event of audits. Clear, tocols yield better results than excessive security meas-
comprehensive and well-documented protocols are invalu- ures that impair workflow and often motivate
able in this effort. New professional profiles, such as Biomed- personnel to find workarounds.
ical Engineers, combining the life sciences and technology - Reliability: When AI is used to drive healthcare deci-
backgrounds required by this kind of endeavor, are increas- sion-making there is an inherent risk of error. More to
ingly available and slowly being integrated into the field. the point, there is a risk of errors being committed in a
Nuclear Medicine is a particularly challenging field in this different manner as a human or a conventional data
sense, due to the overlap of nuclear physics and hybrid imag- processing algorithm would. A human will (in most
ing modalities, each requiring specialized competencies and a cases) acknowledge uncertainty and account for it in
considerable technical background. the decision-making process. Conventional processing
Reference data, such as phantom measurements and can fail more or less graciously, depending on the qual-
healthy volunteer databases can also provide quantitative ity of its programming, but failure will more often than
metrics of data consistency, nowadays easily automated and not be recognizable. That is not necessarily the case
tracked. Reference datasets can also be of great value for with AI, particularly with deep learning models. Fail-
study sponsors trying to determine whether a site is suited ure can sometimes come in unexpected ways, even for
for their study, or in the case that an AI model needs site-spe- apparently minor variations of the model input (an
cific re-training. effect poetically termed “artificial stupidity”). Until suf-
Current examples of AI-related clinical trials in the field ficiently tested in real-world scenarios, AI should be
of Nuclear Medicine are mostly focused on the evaluation considered as complementary to, not independent
of AI clinical tools. These range from the relatively small from, expert decision-making. Notice how, even in
single-site studies (eg, “Symmetricity-Driven Learning that case, unwanted bias may arise if experts find it eas-
Framework for Pediatric Temporal Lobe Epilepsy Detection ier (or safer) to agree with the system rather than hav-
Using 18F-FDG-PET Imaging” NCT04169581; “Usefulness ing to justify their disagreement, hence transforming
of Deep-Learning Image Reconstruction for Cardiac Com- the initial complementary AI in the de-facto decision
puted Tomography Angiography - a Prospective, Non- system.
randomized Observational Trial” NCT03980470), to large - Training: Due to the complementary nature of the role
multisite collaborations (eg, “Automated Quantification of of AI in the immediate future, it is important to train
Coronary Artery Calcifications on Radiotherapy Planning clinical personnel to team up with these new tools. AI
CTs for Cardiovascular Risk Prediction in Breast Cancer does have the potential to identify patterns that elude
Patients: the BRAGATSTON Study” NCT03206333). Evi- human experts and the necessary supervisory role of
dence of the usefulness of AI tools for clinical trial support, the latter cannot come at a cost of overlooking and
on the other hand, is mainly limited to preliminary study overruling the insights that the algorithm can provide.
results. Clinicians will have to strike a balance between trusting
the machine and falling into the complacency of
blindly trusting it.
- Bias: A known example of the limitations of AI algo-
Clinical and ethical rithms is their dependency with the database used to
considerations of AI-enhanced train them. Subtle biases in the database can be inad-
vertently encoded in the resulting model, leading to
trials infamous cases such as algorithms basing their deci-
The ethical implications of the increasing use of AI in health- sions on race, gender, or socioeconomic information.
care are a complex topic still being widely debated.67 Depart- Avoiding these unfortunate situations depends on the
ments entering clinical trials enhanced by AI will, in most same factors that have already been discussed: Suitable
cases, be working with predictive decision-support systems. personnel training to ensure unbiased recruitment,
In which case they may want to consider some of the good record keeping practices, expert supervision of
118 G. Delso et al.
AI output, etc. In those cases where pretrained AI 7. Ervine C: Directive 2004/39/Ec of the European Parliament and of the
models are to be applied, re-training should be consid- Council of 21 April 2004. Core Statutes on Company Law. London:
Macmillan Education UK, 757-759, 2015 http://link.springer.com/
ered, in order to adapt the model to the local popula-
10.1007/978-1-137-54507-7_21
tion (eg, certain tracers are only available in certain 8. EUR-Lex - 32005L0028 - EN - EUR-Lex. https://eur-lex.europa.eu/eli/
countries, and tracer dosing guidelines may vary dir/2005/28/oj
depending on the jurisdiction). 9. Stolberg HO, Norman G, Trop I: Randomized Controlled Trials. AM J
- Explainability: A whole field has appeared recently Roentgenol 183:1539-1544, 2004
around the concept of explainable AI, that is, the devel- 10. Pallmann P, Bedding AW, Choodari-Oskooei B, et al: Adaptive designs
in clinical trials: why use them, and how to run and report them. BMC
opment of AI models capable of providing the reasons Med 16:29, 2018
for its decisions.68 Scientists working in this field work 11. Moore TJ, Zhang H, Anderson G, et al: Estimated Costs of Pivotal Trials
towards more transparent AI, as opposed to the infa- for Novel Therapeutic Agents Approved by the US Food and Drug
mous “black box” behavior of many current models. Administration, 2015-2016. JAMA Intern Med 178:1451-1457, 2018
- Liability: Questions have been raised about the possible 12. Harrer S, Shah P, Antony B, et al: Artificial Intelligence for Clinical Trial
Design. Trends Pharmacol Sci 40:577-591, 2019
legal implications of using AI algorithms of the “black 13. Uribe CF, Mathotaarachchi S, Gaudet V, et al: Machine Learning in
box” type in clinical decision making. It is a concern Nuclear Medicine: Part 1—Introduction. J Nucl Med 60:451-458, 2019
that, in case of malpractice litigation, AI users would 14. Bender E: Big data in biomedicine. Nature 527(7576):S1, 2015
not be able to provide the rationale for the algorithm 15. Stephens ZD, Lee SY, Faghri F, et al: Big Data: Astronomical or Genomi-
cal? PLOS Biol 13:e1002195, 2015
decision. Thus, ultimately the physician (at least cur-
16. Big hopes for big data. Nat Med 26:1, 2020
rently) has the last say in diagnostic as well as thera- 17. Shah P, Kendall F, Khozin S, et al: Artificial intelligence and machine
peutic decision making. learning in clinical development: a translational perspective. Npj Digit
Med 2:1-5, 2019
18. Schneider P, Walters WP, Plowright AT, et al: Rethinking drug design in
the artificial intelligence era. Nat Rev Drug Discov 19:353-364, 2020
Conclusion 19. Nensa F, Demircioglu A, Rischpler C: Artificial Intelligence in Nuclear
Industry stakeholders are looking towards AI hoping to Medicine. J Nucl Med 60(Supplement 2):29S-37S, 2019
20. Wolff J, Pauling J, Keck A, et al: The Economic Impact of Artificial Intel-
break the diminishing returns trend of modern clinical trials. ligence in Health Care: Systematic Review. J Med Internet Res 22:
Modern AI has the potential to provide more meaningful and e16866, 2020
cost-effective trial outcomes. Its applications range from 21. Mak K-K, Pichika MR: Artificial intelligence in drug development: pres-
strategic analysis to compound development, participant ent status and future prospects. Drug Discov Today 24:773-780, 2019
recruitment and clinical decision-making. Establishing col- 22. Reymond J-L, van Deursen R, Blum LC, et al: Chemical space as a
source for new drugs. MedChemComm 1:30-38, 2010
laborations with academic, clinical and technology providers 23. Kadurin A, Nikolenko S, Khrabrov K, et al: druGAN: An Advanced
will be essential to enable these technologies. Generative Adversarial Autoencoder Model for de Novo Generation of
Nuclear medicine departments wishing to enter this field New Molecules with Desired Molecular Properties in Silico. Mol Pharm
will need to undergo several changes, both in terms of tech- 14:3098-3104, 2017
nical infrastructure and in terms of professional practices. 24. Born J, Manica M, Oskooei A, et al: Designing Anticancer Drugs From
Transcriptomic Data via Reinforcement Learning editor. In: Schwartz R
From patient selection to radiotracer production, dosing, (ed): Research in Computational Molecular Biology, Cham: Springer
diagnosis and therapeutic outcome prediction, it will ulti- International Publishing, 231-233, 2020
mately be in the hands of the qualified experts to make sure 25. Boehme P, Hansen A, Roubenoff R, et al: How soon will digital end-
that the right AI tools and the right data are being used, effi- points become a cornerstone for future drug development? Drug Discov
ciently, securely and without biases. Team engagement and Today 24:16-19, 2019
26. Cirillo D, Catuara-Solarz S, Morey C, et al: Sex and gender differences
training will be key factors for these new technologies to take and biases in artificial intelligence for biomedicine and healthcare. Npj
root and yield their full potential. Digit Med 3:1-11, 2020
27. Saltelli A, Bammer G, Bruno I, et al: Five ways to ensure that models
serve society: a manifesto. Nature 582(7813):482-484, 2020
References 28. Feijoo F, Palopoli M, Bernstein J, et al: Key indicators of phase transition
for clinical trials through machine learning. Drug Discov Today 25:414-
1. Scannell JW, Blanckley A, Boldon H, et al: Diagnosing the decline in phar-
421, 2020
maceutical R&D efficiency. Nat Rev Drug Discov 11:191-200, 2012
29. Litster J, Bogle IDL: Smart Process Manufacturing for Formulated Prod-
2. Mullard A: R&D returns continue to fall. Nat Rev Drug Discov 16:9,
ucts. Engineering 5:1003-1009, 2019
2017
30. Wang H, Zhou Z, Li Y, et al: Comparison of machine learning
3. Topol EJ: High-performance medicine: the convergence of human and
methods for classifying mediastinal lymph node metastasis of non-
artificial intelligence. Nat Med 25:44-56, 2019
small cell lung cancer from 18F-FDG PET/CT images. EJNMMI Res
4. Ramagopalan SV, Simpson A, Sammon C: Can real-world data really
7:11, 2017
replace randomised clinical trials? BMC Med 18, 2020. https://www.
31. Hosny A, Parmar C, Quackenbush J, et al: Artificial intelligence in radi-
ncbi.nlm.nih.gov/pmc/articles/PMC6961357/
ology. Nat Rev Cancer 18:500-510, 2018
5. WMA - The World Medical Association-WMA Declaration of Helsinki –
32. Choi H: Deep Learning in Nuclear Medicine and Molecular Imaging:
Ethical Principles for Medical Research Involving Human Subjects.
Current Perspectives and Future Directions. Nucl Med Mol Imaging
https://www.wma.net/policies-post/wma-declaration-of-helsinki-ethical-
52:109-118, 2018
principles-for-medical-research-involving-human-subjects/
33. Liu X, Chen K, Wu T, et al: Use of multimodality imaging and artificial
6. The Medicines for Human Use (Clinical Trials) Regulations 2004.
intelligence for diagnosis and prognosis of early stages of Alzheimer's
Queen's Printer of Acts of Parliament. https://www.legislation.gov.uk/
disease. Transl Res 194:56-67, 2018
uksi/2004/1031/contents/made
How to Design AI-Driven Clinical Trials in Nuclear Medicine 119
34. Hall M: Artificial intelligence and nuclear medicine. Nucl Med Commun 52. Zimmerman BE, Grošev D, Buvat I, et al: Multi-centre evaluation of
40:1-2, 2019 accuracy and reproducibility of planar and SPECT image quantification:
35. Visvikis D, Cheze C, Jaouen V, et al: Artificial intelligence, machine An IAEA phantom study. Z Für Med Phys 27:98-112, 2017
(deep) learning and radio(geno)mics: definitions and nuclear medicine 53. Schwarz SW, Decristoforo C, Goodbody AE, et al: Harmonization of
imaging applications. Eur J Nucl Med Mol Imaging 46:2630-2637, United States, European Union and Canadian First-in-Human Regula-
2019 tory Requirements for Radiopharmaceuticals—Is This Possible? J Nucl
36. Duffy IR, Boyle AJ, Vasdev N: Improving PET Imaging Acquisition and Med 118:209460, 2018. Sep 27jnumed
Analysis With Machine Learning: A Narrative Review With Focus on Alz- 54. Mada MO, Hindmarch P, Stirling J, et al: Competencies and training of
heimer's Disease and Oncology. Mol Imaging 18:1536012119869070, 2019 radiographers and technologists for PET/MR imaging - a study from the
37. Tseng H-H, Wei L, Cui S, et al: Machine Learning and Imaging Infor- UK MR-PET network. Eur J Hybrid Imaging 4:1, 2020
matics in Oncology. Oncology 98:344-362, 2020 55. Zhang J, Jacko R, Vu T, et al: Artificial Intelligence Integration in Cloud-
38. Hatt M, Laurent B, Ouahabi A, et al: The first MICCAI challenge on PET based Real-time Data Quality Assurance for Multi-Institutional Clinical
tumor segmentation. Med Image Anal 44:177-195, 2018 Trials. J Nucl Med 59(supplement 1):1280, 2018
39. Xiang L, Qiao Y, Nie D, et al: Deep Auto-context Convolutional Neural 56. Zou D, Ma L, Yu J, et al: Biological Databases for Human Research.
Networks for Standard-Dose PET Image Estimation from Low-Dose Genomics Proteomics Bioinformatics 13:55-63, 2015
PET/MRI. Neurocomputing 267:406-416, 2017 57. Global Alliance for Genomics and Health. GENOMICS. A federated eco-
40. Kim K, Wu D, Gong K, et al: Penalized PET reconstruction using deep system for sharing genomic, clinical data. Science 352(6291):1278-
learning prior and local linear fitting. IEEE Trans Med Imaging 1280, 2016
37:1478-1487, 2018 58. Blomberg N, Consortium E: ELIXIR position paper on FAIR data man-
41. Häggström I, Schmidtlein CR, Campanella G, et al: DeepPET: A deep agement in the life sciences. F1000 Research 6, 2017. https://
encoder-decoder network for directly solving the PET image reconstruc- f1000research.com/documents/6-1857
tion inverse problem. Med Image Anal 54:253-262, 2019 59. The Alliance of Genome Resources: Building a Modern Data Ecosystem
42. Liu F, Jang H, Kijowski R, et al: Deep Learning MR Imaging-based Atten- for Model Organism Databases. Genetics 213:1189-1196, 2019
uation Correction for PET/MR Imaging. Radiology 286:676-684, 2018 60. Wilkinson MD, Dumontier M, Aalbersberg I, et al: The FAIR Guiding
43. Kang SK, Seo S, Shin SA, et al: Adaptive template generation for amyloid Principles for scientific data management and stewardship. Sci Data 3,
PET using a deep learning approach. Hum Brain Mapp 39:3769-3778, 2016. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4792175/
2018 61. Linden M, Prochazka M, Lappalainen I, et al: Common ELIXIR Service
44. Suk H-I, Lee S-W, Shen D: Latent feature representation with stacked auto- for Researcher Authentication and Authorisation. F1000 Research: 7,
encoder for AD/MCI diagnosis. Brain Struct Funct 220:841-859, 2015 2018. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6124379/
45. Huang C, Clayton EA, Matyunina LV, et al: Machine learning predicts 62. Bowles KH, Potashnik S, Ratcliffe SJ, et al: Conducting Research Using
individual cancer patient responses to therapeutic drugs with high accu- the Electronic Health Record Across Multi-Hospital Systems: Semantic
racy. Sci Rep 8:16444, 2018 Harmonization Implications for Administrators. J Nurs Adm 43:355-
46. Weiss J, Hoffmann U, Aerts HJWL: Artificial intelligence-derived imag- 360, 2013
ing biomarkers to improve population health. Lancet Digit Health 2: 63. Geneviève LD, Martani A, Mallet MC, et al: Factors influencing harmo-
e154-e155, 2020 nized health data collection, sharing and linkage in Denmark and Swit-
47. MachinE Learning Ledger Orchestration for Drug DiscoverY | MEL- zerland: A systematic review. PLOS ONE 14:e0226015, 2019
LODDY Project | H2020 | CORDIS | European Commission. https://cor- 64. Dondi M, Kashyap R, Pascual T, et al: Quality Management in Nuclear
dis.europa.eu/project/id/831472 Medicine for Better Patient Care: The IAEA Program. Semin Nucl Med
48. Makris NE, Huisman MC, Kinahan PE, et al: Evaluation of strategies 43:167-171, 2013
towards harmonization of FDG PET/CT studies in multicentre trials: 65. Dondi M, Torres L, Marengo M, et al: Comprehensive Auditing in
comparison of scanner validation phantoms and data analysis proce- Nuclear Medicine Through the International Atomic Energy Agency
dures. Eur J Nucl Med Mol Imaging 40:1507-1515, 2013 Quality Management Audits in Nuclear Medicine (QUANUM) Program.
49. Treves ST, Lassmann M: International guidelines for pediatric radio- Part 1: the QUANUM Program and Methodology. Semin Nucl Med
pharmaceutical administered activities. J Nucl Med 55:869-870, 2014 47:680-686, 2017
50. Buvat I, Frey E, Green A, et al: Quantitative Nuclear Medicine Imaging: 66. Frangi AF, Tsaftaris SA, Prince JL: Simulation and Synthesis in Medical
Concepts. Requirements and Methods. Hum Health Rep: 9, 2014. Imaging. IEEE Trans Med Imaging 37:673-679, 2018
http://lup.lub.lu.se/record/4390492 67. Rigby MJ: Ethical Dimensions of Using Artificial Intelligence in Health
51. Boellaard R, Delgado-Bolton R, Oyen WJ, et al: FDG PET/CT: EANM Care. AMA J Ethics 21:121-124, 2019
procedure guidelines for tumour PET imaging: version 2.0. Eur J Nucl 68. Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, et al: Explainable Artifi-
Med Mol Imaging: 42, 2015. https://doi.org/10.1007/s00259-014- cial Intelligence (XAI): Concepts, taxonomies, opportunities and chal-
2961-x lenges toward responsible AI. Inf Fusion 58:82-115, 2020