Focus issue: Artificial intelligence in medical physics

Abstract

Following on the European Federation of Organisations for Medical Physics (EFOMP) editorial on Artificial Intelligence in relation to the medical physics profession [[1]Kortesniemi M. Tsapaki V. Trianni A. Russo P. Maas A. Källman H.-E. et al.The European Federation of Organisations for Medical Physics (EFOMP) white paper: big data and deep learning in medical imaging and in relation to medical physics profession.Phys Med. 2018; 56: 90-93https://doi.org/10.1016/j.ejmp.2018.11.005Abstract Full Text Full Text PDF PubMed Scopus (23) Google Scholar], and in order to meet the educational needs of the Medical Physicist (MP) in this new area of AI, EFOMP announced in June 2019 the creation of a 2 years Working Group (WG) entitled “Artificial Intelligence (AI)”. The expected outcomes are an AI for Medical Physicists (MPs) Curricular and Professional Program as well as an EFOMP European School of Medical Physics Expert (ESMPE) AI module. EFOMP identified the need for Medical physicists (MPs) to take up their stakeholders’ role in the era of AI in medicine, by updating their training and education programs. This is also clearly stated in the EU RP174 presenting European Guidelines on Medical Physics Expert and strengthening the importance of education and training as the foundations of the MP profession [[2]EC. RP174 European guidelines on the medical physics expert; 2014. http://op.europa.eu/en/publication-detail/-/publication/b82ed768-4c50-4c9a-a789-98a3b0df5391 [accessed January 29, 2021].Google Scholar]. It is within this context that the EFOMP WG on AI proposed to the Editor-in-Chief of EJMP a focus issue (FI) dedicated to AI in medical physics. The idea was to gather in one Physica Medica volume the most important topics addressed in the curriculum, for the education and training of European Medical Physicists (MPs). Having received the enthusiastic approval of the Editor, who suggested to widen the scope of this Focus Issue to current research aspects in the field of AI, we were asked to act as Guest Editors. We were honoured and humbly accepted the task. This FI aims at providing a summary of the techniques and applications of AI in medical physics. It also addresses common pitfalls associated with these technologies. Because the application of AI in medicine in general and in medical physics in particular has seen an unprecedented increase in the recent years, the medical physicist profession has to keep pace with these changes, and we hope that this Focus Issue will provide a guide for MPs who are, or will be involved, in this exciting field. Twenty-seven papers (out of the 32 invited) were finally accepted for this Focus Issue. The table of content of this issue is meant to introducing MPs, through review and original papers, to the pillars of knowledge, development and applications of AI, in the context of medical imaging and radiation therapy. It is certainly not exhaustive, and it lacks some important topics like ethical aspects of AI or a methodological paper on the integration of AI applications in the clinical workflow, which unfortunately did not make it through the peer-reviewed process of the invited contributions. Because of the restrictive timeframe for the FI there was no possibility to include new contributions on these topics. As the curricular program developed by the WG was almost ready at the time the FI was conceived, it was also included as an invited (and accepted) paper. The FI starts therefore with it and the following articles are presented in a sequential way that reflects the curricular structure. The curricular programs for MPs are organized per sub-specialties. The curriculum on AI developed by the working group (Zanca et al.) is not meant to replace them, but rather expand the sub-specialties’ ones on topics related to AI [[3]Zanca F. Hernandez-Giron I. Avanzo M. Guidi G. Crijns W. Diaz O. et al.Expanding the medical physicist curricular and professional programme to include artificial intelligence.Phys Med. 2021; 83: 174-183https://doi.org/10.1016/j.ejmp.2021.01.069Abstract Full Text Full Text PDF PubMed Scopus (15) Google Scholar]. It has been subdivided in two levels, Basic and Advanced, depending on the potential involvement of the medical physicist in specific applications of AI and hence allowing for tailored education. The aim of this Basic training level is to introduce MPs to the pillars of knowledge, development and applications of AI, in the context of medical imaging and radiotherapy. The Advanced level instead, aims at building deeper expertise in the same topics. This paper is an extremely important contribution to the education of medical physicists in AI as the curriculum presented could be used not only as a module of the ESMP but also as a guideline for the programs of national organisations, members of EFOMP. Note that surveys had been performed to assess the perceptions of MPs towards relevance and impact of AI [[4]Diaz O. Guidi G. Ivashchenko O. Colgan N. Zanca F. Artificial intelligence in the medical physics community: an international survey.Phys Med. 2021; 81: 141-146https://doi.org/10.1016/j.ejmp.2020.11.037Abstract Full Text Full Text PDF PubMed Scopus (13) Google Scholar]. Avanzo et al. [[5]Avanzo M. Porzio M. Lorenzon L. Milan L. Sghedoni R. Russo G. Massafra R. Fanizzi A. Barucci A. Ardu V. Branchini M. Giannelli M. Gallio E. Cilla S. Tangaro S. Lombardi A. Pirrone G. De Martin E. Giuliano A. Belmonte G. Russo S. Rampado O. Mettivier G. Artificial intelligence applications in medical imaging: A review of the medical physics research in Italy.Phys. Med. 2021; 83 (Online ahead of print.PMID: 33951590): 221-241https://doi.org/10.1016/j.ejmp.2021.04.010Abstract Full Text Full Text PDF PubMed Scopus (19) Google Scholar] presented a survey of research articles in AI applications in medical imaging produced in 2015–2020 by authors with scientific affiliation in Italy, also in collaboration with the task group “AI for Medical Physics” of the Italian Association of Medical Physics (AIFM). This represents the first attempt to review this research field in Italy by the medical physics community, with an analysis of 168 studies. The vast majority (71%) was in the field of diagnostic imaging (MRI, CT, radiography, mammography), and prevalently aiming at image classification tasks (57% of the articles) and then at image segmentation (16%), using deep learning in only 35% of the cases. These findings confirmed for the Italian framework the well-known rapid growth of the research interest in AI technologies by the international medical physics community in very recent years. They also pointed to possible difficulties in assembling and accessing large databases of images best suited for analysis by deep learning AI methods. This concern was also shared at EU level with specific research calls dedicated to building large repositories of freely available medical images for AI applications. In Montero et al. [[6]Barragán-Montero Ana Javaid Umair Valdés Gilmer Nguyen Dan Desbordes Paul Macq Benoit Willems Siri Vandewinckele Liesbeth Holmström Mats Löfman Fredrik Michiels Steven Souris Kevin Sterpin Edmond Lee John A. Artificial intelligence and machine learning for medical imaging: A technology review.Eur. J. Med. Phys. 2021; 83: P242-P256https://doi.org/10.1016/j.ejmp.2021.04.016Abstract Full Text Full Text PDF Scopus (22) Google Scholar] a review of the building blocks of AI methods, together with their application to medical imaging is given. A key section is the one describing the state-of-the-art of AI methods such as Machine Learning (Ml) and Deep Learning (DL) for medical imaging analysis, completed by interesting interpretation on the evolution needed for having such AI applications really breaking through in clinical practice. An overview of such ML-based applications in the literature is given by Strigari et al. [[7]Manco L. Maffei N. Strolin S. Vichi S. Bottazzi L. Strigari L. Basic of machine learning and deep learning in imaging for medical physicists.Phys Med. 2021; 83: 194-205https://doi.org/10.1016/j.ejmp.2021.03.026Abstract Full Text Full Text PDF PubMed Scopus (11) Google Scholar]. The manuscript highlights more than 188 scientific papers and discusses some limitations and opportunities of AI application in the clinical practice for future research. The authors identified common factors such as research area, discipline, number of images reported in the study for validation, number of patients analysed, type of images, codes and algorithms used, primary and secondary purposes of the studies. Regardless of the field of work, the reader can easily find and identify his/her own study discipline and have an immediate overview of the state of the art of the literature of the last 5 years on which to identify his/her research and amount of data needed in the field of medical imaging applied to AI. When it comes to developing such applications, as reviewed by Castiglioni et al. [[8]Castiglioni I. Rundo L. Codari M. Di Leo G. Salvatore C. Interlenghi M. et al.AI applications to medical images: from machine learning to deep learning.Phys Med. 2021; 83: 9-24https://doi.org/10.1016/j.ejmp.2021.02.006Abstract Full Text Full Text PDF PubMed Scopus (63) Google Scholar], each of the phases required for building them has its specific challenge. Researchers in AI need to collect a large set of high quality labelled and annotated data, as the accuracy of AI tools depends largely on the dataset used for training. Harmonization techniques can minimize error due to heterogeneity when dealing with data from multiple centers, for instance. In radiomics studies, overfitting of machine learning can be prevented by careful feature selection before ML. Deep neural networks composed stacks of layers of nonlinear units are more challenging due to the infinite possibilities of arranging neurons into different architectures. Methods to estimate sample size in AI are still under investigation, however, data augmentation can deal with small and imbalanced datasets, and transfer learning – the use of pretrained AI tools adapted to the task at hand – can be applied as a less demanding alternative to training a DL network from scratch. To illustrate the aspect of data preparation, a comprehensive guide to open access platforms and tools has been described by Diaz et al. [[9]Diaz O. Kushibar K. Osuala R. Linardos A. Garrucho L. Igual L. et al.Data preparation for artificial intelligence in medical imaging: a comprehensive guide to open-access platforms and tools.Phys Med. 2021; 83: 25-37https://doi.org/10.1016/j.ejmp.2021.02.007Abstract Full Text Full Text PDF PubMed Scopus (20) Google Scholar]. They explain in detail a typical medical image pipeline, i.e. de-identification, data curation, centralised and decentralised medica image storage, and data annotation tools. They provide a comprehensive guide to choose among the armamentarium of currently available tools and platforms towards developing and or applying AI algorithms. Next to in-house developed AI application, commercially available ones are also increasing and dealing with the procurement, commissioning and integration in clinical workflow of such tools poses important challenges. Bosmans et al. [[10]Bosmans Hilde Zanca Federica Gelaude Frederik Procurement, commissioning and QA of AI based solutions: An MPE’s perspective on introducing AI in clinical practice.Eur. J. Med. Phys. 2021; 83: P257-P263https://doi.org/10.1016/j.ejmp.2021.04.006Abstract Full Text Full Text PDF Scopus (7) Google Scholar] proposed a framework to facilitate medical physicists’ role in the introduction of AI solutions in clinical practice. Focus was given to the procurement process including acceptance, commissioning and QA of AI tools. On the AI regime, these steps require further consideration and dedicated test methods as compared to the tradition radiological equipment procurement process. The nature of AI based tools warrant specific Key Performance Indicators (KPIs) and metrics defined for systematic set of clinical cases (in acceptance) and ensuring suitability to local clinical environment (in commissioning). Insight of the expected performance of new clinical AI tools may also be surveyed from scientific publications or published data-sets with representative data, with similarity to local workflow. Quality assurance of AI tools is needed to ensure the stable performance of these algorithms especially concerning the upscaling usage and upgrading with self-learning networks. All these aspects indicate new and additional challenges which must be taken into account while considering continuous professional development of medical physicist and our role in hospitals. Both for in-house developed and for commercially available applications, an effective regulation is crucial to enable a safe and optimal embedding of AI-based medical devices in the clinical settings. As described in Beckers et al. [[11]Beckers R. Kwade Z. Zanca F. The EU medical device regulation: Implications for artificial intelligence-based medical device software in medical physics.Phys Med. 2021; 83: 1-8https://doi.org/10.1016/j.ejmp.2021.02.011Abstract Full Text Full Text PDF PubMed Scopus (18) Google Scholar], in May 2021, the European Medical Device Regulation (EU MDR) will become fully applicable and AI-devices with an impact on patient diagnosis or treatment fall under such directive as well, when classified as medical devices. The paper summarizes the new regulatory roadmap comprising the intended use, risk classification, clinical evidence generation and post-market surveillance. With such knowledge, MPEs will be able to effectively participate on the purchase, commission and introduction of AI-based tools in the clinical workflow. As noted in the paper, one of the main pitfalls at present is the of very guidelines for commission and of AI based medical device the for from national and international medical physicists’ professional to this but not the review of et al. S. John M. imaging and big A review from a medical physics Med. 2021; 83 Full Text Full Text PDF PubMed Scopus Google Scholar] into recent imaging solutions applied to medical physics and settings. The in big data has up opportunities for the application of imaging solutions to big data of The review summarizes the key tools and of imaging and big data in clinical practice and has a of the steps required to a based are into solutions must patient workflow and the AI The key challenges of and data is also in with clinical and of the The use of AI can patient and but the of and must be with training and continuous professional development of medical physicists to ensure of patients and regulatory The integration of AI in and for and diagnosis of These new AI applications, however, require validation, as their and are for patient In et al. N. N. K. J. Performance of an artificial intelligence with clinical workflow integration – of and Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar], a AI with clinical workflow integration was for diagnosis of and on a two of and of patients The AI was able to accuracy for and for potential to and in clinical with AI and diagnosis in CT, the most common for is of interest for In et al. M. I. C. classification in with Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] an AI to in as or which a neural network to image with a machine learning was The were using training. The were in a available and in high performance under the in validation, which the classification performance of the Quality assurance of AI tools in imaging on the use of specific for in and et al. S. F. L. V. R. et in images by deep learning a Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] a with of and of different for AI in They deep learning based on for or and of performance on from potential for patient by AI-based image quality networks are a AI which et al. S. M. M. et networks to image from a Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] the use of for a from a of The images provide of compared with the in of to and similarity images by can diagnostic The applications of AI in medicine on and imaging were reviewed by et al. A. A. I. The of artificial intelligence and deep learning in and Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar], with on deep learning (DL) used on a of applications. They presented DL algorithms and used in and and imaging for different purposes such as image image and diagnosis and as well as radiomics and The AI-based radiation in using and is also the AI are presented from the of of the MP and on their and limitations as tools. As imaging is the most common medical AI is expected to have impact in this field. et al. K. A. imaging quality Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] the use of to diagnostic quality different image quality on patient images patient and The under the were for the and for the and the an is proposed as an quality of which could image and patient AI-based techniques have to be important tools not only for but also for MPs and from medical imaging is among most from is as of and use of deep neural networks et al. and using a deep learning patients J. Med. Phys. 2021; 83: Full Text Full Text PDF Scopus Google Scholar] that and could be The was by and tools available in the The developed tools and databases used, are available for the Medical Physics community by the et al. R. Mettivier G. M. A. G. S. et deep learning for Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] a for the classification of the of in based on a deep neural network were compared with state-of-the-art of and on two different A was also by the in the image at risk of of a can facilitate This could be further to also include of and within the With its is key for the and treatment of and AI applications to imaging are rapid pace of In the review of et al. J. V. A. Machine learning in imaging: image Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar], the authors the clinical applications, and current of used in on AI for image identify current machine learning used to the image in clinical images in order to up is an a process. The image quality and its potential for into clinical use are In a clinical quality of images is the role and of AI methods are expected to diagnosis in with impact on and et al. J. A. W. et applications of in Med. 2021; 83: Full Text Full Text PDF PubMed Scopus (11) Google Scholar] review the for AI applications to fully data for classification or patient and identify the The original research papers reviewed by the authors were classified based on their applications, into diagnosis and and the applications from the classification, and The authors identified the and potential of applications, the and the for of the of technologies in clinical practice. AI application with clinical impact in is presented by et al. P. R. C. C. et and in a deep learning with Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar], deep learning for the and of and The neural networks of for and of the in an of a high performance in and is for AI applications, to and up the workflow from patient to treatment and et al. N. for using and deep learning Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] deep learning applications in performed in a They a to the optimal that the from a treatment has been to meet the The can be used as a for et al. M. G. Michiels S. K. Lee et learning for of the of data quality and on Med. 2021; 83: Full Text Full Text PDF PubMed Scopus (13) Google Scholar] the factors deep learning for of They the of different and and of that accuracy depends on data size and This is a typical of AI is currently in the of treatment is to and diagnostic data from CT, and in and AI applications and methods which have been developed and in described are also and the which can be seen as a for AI patient et al. J. N. L. et deep classification in images for Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] use a DL to on clinical proposed in high accuracy for for compared to clinical allowing and diagnosis of the of the the This study is a towards of for et al. P. L. N. W. F. L. et deep learning in radiomics and challenges of and data Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] the of the research on AI for image from medical images, a field currently The authors review the methods that the AI for classification and in clinical practice. They also the crucial of of large datasets, increasing the as to the potential clinical of radiomics and the development of AI In Maffei et al. M. et al.AI on radiomics to segmentation quality of for Med. 2021; 83: Full Text Full Text PDF PubMed Scopus Google Scholar] a radiomics was to segmentation quality of to radiomics were to be and the AI and with an accuracy of The proposed workflow an of segmentation quality and may the of an as well as of large treatment This FI of Physica Medica reflects the interest of the scientific community and the professional by EFOMP in AI by the techniques and the particular applications of AI in medical physics as well as the current limitations and the challenges to The need for education or continuous professional development of medical physicists to with the use of AI solutions in medical physics is also the presented in this FI of high interest for the of Physica

Ähnliche Arbeiten