AI in Medical Imaging: A Growing Field

Background and rationale

When Doi (2007) wrote on the expected successes in computer-aided diagnostics popularly referred to as CAD, he probably underestimated the thrust of artificial intelligence which he referred to as automated computer diagnostic. Today artificial intelligence (AI) is (completely) taking over almost every profession and being termed as disruptive technology. The medical field has not been spared. Whereas AI was first coined in mid-20th century by Arthur Samuel its prevalence has only been recently felt. Basically, AI refers to the ability of computerized device ability to read and learn thus performing the basic human-brain skills (Pesapane et al., 2018). For such device to function it requires to be trained with data in its sphere of use. The prevalence of AI has been occasioned with the invention of deep learning techniques which basically introduces non-linear representation-learning methods to process input data, classify it and produce diagnostic treatment (Shiraishi et al., 2011). For AI to be extremely effective using DL, the data is processed through algorithms that work similarly to the human brain’s neural networks (Pesapane et al., 2018). Usually medical imaging technologists are faced with immense data that takes a lot of time to process and make deductions hence limiting the time taken for other clinical treatment measures (Jha & Topol, 2016). AI reduces this amount of time spent on diagnosis and interpretation. This has come with its own criticisms and questions of whether AI is a job-replacing technology. The debate has also extended to the ethical implications of having AI-led diagnosis especially in terms of bias.

The field of medical imaging has been at the forefront of digitization since the 20th century but questions of how far AI should be incorporated in the field have gained traction in the last decade than before. Already the implications of AI in medical imaging particularly in chest radiography, ultra-sound and computed tomography scan have been the source of increased publications which in the last decade have increased to approximately 700 yearly (Pesapane et al., 2018). A study conducted by Codari et al., (2019) demonstrated that radiologists predicted the impacts of AI but it was not certain whether the impacts were felt as an increase or decrease in the roles the radiologists played. This is shown in the figure 1. below.

When Doi (2007) wrote on the expected successes in computer-aided diagnostics popularly referred to as CAD, he probably underestimated the thrust of artificial intelligence which he referred to as automated computer diagnostic. Today artificial intelligence (AI) is (completely) taking over almost every profession and being termed as disruptive technology. The medical field has not been spared. Whereas AI was first coined in mid-20th century by Arthur Samuel its prevalence has only been recently felt. Basically, AI refers to the ability of computerized device ability to read and learn thus performing the basic human-brain skills (Pesapane et al., 2018). For such device to function it requires to be trained with data in its sphere of use. The prevalence of AI has been occasioned with the invention of deep learning techniques which basically introduces non-linear representation-learning methods to process input data, classify it and produce diagnostic treatment (Shiraishi et al., 2011). For AI to be extremely effective using DL, the data is processed through algorithms that work similarly to the human brain’s neural networks (Pesapane et al., 2018). Usually medical imaging technologists are faced with immense data that takes a lot of time to process and make deductions hence limiting the time taken for other clinical treatment measures (Jha & Topol, 2016). AI reduces this amount of time spent on diagnosis and interpretation. This has come with its own criticisms and questions of whether AI is a job-replacing technology. The debate has also extended to the ethical implications of having AI-led diagnosis especially in terms of bias. The field of medical imaging has been at the forefront of digitization since the 20th century but questions of how far AI should be incorporated in the field have gained traction in the last decade than before. Already the implications of AI in medical imaging particularly in chest radiography, ultra-sound and computed tomography scan have been the source of increased publications which in the last decade have increased to approximately 700 yearly (Pesapane et al., 2018). A study conducted by Codari et al., (2019) demonstrated that radiologists predicted the impacts of AI but it was not certain whether the impacts were felt as an increase or decrease in the roles the radiologists played. This is shown in the figure 1. below. Figure 1. Represents the number of medical respondents who predicted the impacts of AI in medical imaging. Source: Codari et al., (2019) The impacts of AI particularly for radiographers are expected to be extensive as deep learning displaces their role in the imaging process by performing the wholesome task. Radiologists may still have an important role to play in the diagnosis process unlike radiographers (Topol, 2019). Hence AI’s role in radiography needs immediate analysis which this research offers. This research, therefore, examines whether and how AI efficacy corroborates the performance of radiographers and radiologists without necessarily replacing human radiographers and radiologists. Currently, because of the high pace in which AI has been modifying and changing, there are no clear guidelines on its use. The only ambitious policy initiative being the European Parliament’s guidelines recommended to the Commission on Civil Law Rules on Robotics which only speak to the ethical issues of privacy of patient data used to train the AI (Neri et al., 2020).

Review of literature and identification of current gap in knowledge

Whereas there is sufficient literature on the advent of AI in medical imaging, most of the published articles have paid too much attention on its impact in radiology as a profession particularly the diagnostic capabilities of AI. The space of radiographers has been diminished within those articles hence making deductions on the impact of AI on radiographers is an extremely strenuous task. Perhaps looking at AI from a medical imaging point of view gives more insight. Recent studies have shown that there is a rapid increase in the workload of medical imaging technologists. Indeed, as Alexander et al., (2020) connotes, in a study done comprising 50 radiologists in the United States of America (USA), almost all recorded that their daily tasks had increased with 14 of them reporting a 20% increase in their tasks. This therefore underscores the analysis by Jha and Topol (2016) that the increased workload results to wastage of time in image analysis as well as increased fatigue which leads to poor diagnosis and quality imaging (Pesapane, Codari and Sardanelli, 2018). Ironically, the increased development of AI and the increased tasks for medical imaging technologists appears not to resonate with the desired outcome which is an increase in AI resulting to less workload for medical imaging technologists. A further study done by Alexander et al., (2020) shows that most radiologists have expressed mistrust on the full automation of medical imaging occasioned by lack of regulatory approval within the USA. However, this study fails to depict the import of AI on radiographers and their opinion on the use of AI devices in their sphere as most radiologists would be scared to allow a device to diagnose their patients.

All these underscores the importance of developing deep learning tools for AI that helps in reducing workloads as it works without any further physical input of the radiographer in taking and processing images (King, 2018). Shiraishi et al., (2011), in a study shows that deep learning tool incorporated to detect solid tumour in the same manner as the standard form used, was not affected by differences in interpretation yet it yielded the proper diagnosis accepted by radiologists. This means that segmentation of organs need not be instructed on the device by the radiographer during the imaging but the AI tool automatically segments the body allowing easier analysis of the image (Lewis et al., 2019) with evidence suggesting that Australian medical imaging technologists have championed the application for such tools resulting to efficient delivery of imaging services. However, consumed time cannot be saved in the imaging process without proper training of AI devices. Indeed, the scepticism registered by Alexander et al., 2020, is because of lack of proper training of devices. This position is galvanized by Hosny et al., 2018 who upholds that the biggest challenge today is having AI systems that can only perform one job which they attribute to the lack of intensive training of these systems. The introduction of deep learning especially on methods such as convolutional neural networks (CNN) allows the AI to consume big data during training thus increasing the efficiency and the output capabilities of these AI systems (Pesapane et al., 2018). Accordingly, this enables the AI to conduct searches on cloud-based databases and other electronic records for efficient characterisation of images (Kohli et al., 2017). Moreover, Kohli et al., (2017) suggests that deep learning allows the AI to output images that clearly depict features in the area of concentration that are can or cannot be seen with human eyes. This means that the productivity of radiographers is improved as they reduce the time taken by radiologists to interpret diagnosis given by AI.

Shiraishi et al., (2011) furthers this argument by suggesting that the use of artificial neural network (ANN), a processing mechanism in computer-aided diagnosis (CAD), allows the radiographer to incorporate different imaging parameters in a single scan hence reducing the imaging period. However, in as much as this analysis is relevant it only relates to CADs which cannot match the advances in AI since AI incorporates deeper ANNs such as CNNs. This however means that AI would perform much better than CAD. Safdara et al., (2020) cautions against the blanket use of AI in medical imaging particularly chest radiography which has many rare diseases that AI may not be trained on yet thus resulting to high numbers of false positives and false negatives which may cause fatal misdiagnosis. Moreover, as training of machines is usually based on the majority population available at the place where it is developed, the training data may be dissimilar with the place of application in terms of patient characteristics thus resulting to misdiagnosis (Park & Han, 2018; Safdara et al., 2020). Yet this is a more reason to enhance training of these AI during their application so as to enable them increase their productivity and avoid the low intelligence which Hosny et al (2018) raises. But this raises a further question, of which many studies have raised including Pesapane et al., (2018), on whether AI in medical imaging will replace human radiographers and by extension radiologists. Jha & Topol (2018), suggest that AI will not “replace” but rather “displace” medical imaging technologists. Furthermore Shen et al., (2017), opines that deep learning tools in AI would allow automation of most of the preparatory measures that radiographers have to implement for proper segmentation and imaging of the desired area of concern. In as much as Hosny et al., (2018) suggest this to be an important milestone of AI, the human radiographers have a reason to fear that their jobs would be completely taken over by AI. On the contrary, Pesapane et al., (2018), relying on Abraham Lincoln’s words, suggests that medical imaging technologists should embrace AI so as to be part of the future; to predict the future is to create it. They suggest that in as much as AI will take over some of the routine tasks, their role is important especially in relaying of AI-diagnosis information to patients as AI is not empathic. They conclude that their roles will be restructured to fit in other roles within the medical imaging arena. They however caution that for these radiologists to manage not being replaced them must accept to train themselves on the ways in which they can help in building efficiency using AI. However, their analysis is too focused on the use of AI in diagnosis of patients thus alienating the impact of AI on a radiographer as radiographers are not usually involved in the interpretation of images.

According to Lewis, Gandomkar & Brennan (2019), suggested that radiographers will not be left out in the process of introducing AIs as they have a very significant duty to perform in continuous training of these AIs with high quality images and feeding them with data for the numerous diseases and manners that they exhibit themselves. Moreover, it was opined that AIs need to be continually optimised to eliminate problems of bias and increase their intelligence. That even for radiographers, AIs need to be continuously trained on the different imaging protocols in the various imaging methods and particularly ultrasound, CT scan among others. Already it has been noted that current AIs may not have the capacity to detect certain diseases which further underscores the important role of radiographers (Safdara et al., 2020). Codari et al., (2019) studied that a large number of medics practised image modalities but fewer of them especially in radiography, ultrasound, angiography/fluoroscopy and hybrid imaging believed that the modalities could be used to develop AI in the future as shown in figure 3.

However, Lewis et al., (2019) opined that radiography curriculums need to incorporate AI as a core subject so as to prepare the students for a working environment with AI. These skills will be extremely important to enable them note errors which these devices may making due to poor or intended wrongful deep learning. Neri et al., (2020) argued that only those technologists with background knowledge on AI imaging should be relied upon in developing deep learning tools and in validating AI imaging and diagnosis. However, these studies and researches fail to take into consideration the dynamism that comes with AI which even regulatory authorities have found difficult to legislate and establish governing rules. This is caused by the fact that AIs will keep on learning and rectifying mistakes almost instantaneously thus leaving no room for detecting damning ethical errors. Implementation of AI has raised ethical issues even in other fields such as security in relation to face recognition cameras that cannot detect black people (Safdara et al., 2020). Safdara and colleagues strongly raise the question of the implication having an AI device trained on a population whose background is the opposite of its application thus affecting the characteristic diagnosis given. Pesapane et al., (2018) raised concerns of racial extremism being built in those algorithms which data is trained upon. An American-European report suggests prevalence of AI may result to radiographers and radiologists fail to detect algorithmic errors made by AI due to dependency (Geis et al., 2019). For this reason, some scholars suggest that AI in medical imaging should not be allowed to side-line the ability of human radiographers and radiologists to make independent conclusions or enhance protocols of these AI-driven devices (Neri et al., 2020).

Methodology – Justify the proposed methodology

This research proceeds on a qualitative approach of data and studies available in relation to AI and medical imaging. It also takes an intra-professional and inter-professional approach (Adams and Smith, 2003) seeking to ask and attempt to answer questions on what is the future for radiographers in the age of advanced AI. The inter-professional approach will entail research on published articles on the import of AI in radiology which is a related field within the larger medical imaging fraternity.

Theoretical framework

There is adoption of both epistemological and interpretive approach. Interpretivist approach entails the study of issues with the intention of understanding the social implication or from the recipient’s point of view (Pesapane et al., 2018). This research interrogates the implication of the prevalence of AI in medical imaging on radiographers and by extension the radiologist. Epistemology theory in this case entails the study of how AI as a branch of knowledge interacts with medical imaging and its implications on those involved in medical imaging particularly radiographers.

Literature Review

This study utilises a systematic literature review to review existing literature regarding AI in medical imaging. A systematic Literature Review (SLR) aims at identification, integration and critical evaluation of relevant studies that address a particular research question (Siddaway, 2014). The search is exhaustive and transparent and is done over numerous databases and grey literature. This requires having a well laid out search strategy which focuses on answering the question of interest (Gough et al., 2012.). The search strategies, key words and search limits need to be applied in the review (McNally, 2016) The SLR has important principles: focus, transparency, accessibility, integration, equality and coverage. It has the ability to provide answers to broader research questions with minimal bias making it sit at the top of the ‘hierarchy of evidence’ (Siddaway, 2014) therefore, the most suitable method for this study.

Methods and Data

Identifying the research question

Having the right and clear research question is an important part of systematic literature review (Aveyard et al., 2014). This study compared two models of identifying a research question: SPIDER and PEO. PEO is a framework that utilised in a qualitative research question. It examines the impact of exposure to a certain condition or phenomenon (Speckman & Friedly, 2019). The concepts identified in PEO are: Population, Exposure and Outcome(s). SPIDER, on the other hand is a modification of PICO utilised for qualitative/mixed method research and the concepts are: Sample, Phenomenon of Interest, Design, Evaluation and Research type (Methley et al., 2014).Since this study seeks to evaluate the impacts of AI in medical imaging, PEO was considered as the convenient tool and is shown in table 1.

Search strategy

The main search strategy applied for this research is the use of primary scientific databases; selectively and customized random search on Google. All articles relied upon in this research are limited to only those published between 2006 to 2020. There will be the use of scientific databases such as: CINAHL Plus, Medline, PubMed and Cochrane for most of its research. Further research will be done on the National Institutes of Health (NIH) database to generate work that is relevant to this study.

Inclusion and Exclusion Criteria

Work that will be included will consist of articles published in English in the period of 2006 to 2020. This period ensures that information received is not outdated. Any work that will fall off the duration mentioned or be a repetition of another and/or not be answer the specified objectives will be excluded. This criteria of inclusion and exclusion enables scientific rigor of the research and ensures that only relevant information is extracted.

Critical appraisal of the quality of the studies

Critical appraisal is the process of assigning merit to, or evaluating the quality of articles by careful scrutinization. The tool used contains particular criteria in which the studies are compared against to determine the trustworthiness of the results (Barlett et al., 2015). This study will utilise the Critical Appraisal Skills Programme (CASP) to assess the quality of articles obtained from the search. The CASP contains a set of critical appraisal tools that are customized to suit various researches including systematic reviews (Nadelson & Nadelson, 2014)

Ethical issues

This research raises no external ethical issues as it does not entail any interviews or clinical trials. The purpose of this research is to analyse documents and literature on AI and its impact on medical imaging. The only ethical issues concerned are internal issues on how existing literature has approached ethical issues relating to AI.

Resource implications

The use of the electronic search engines to obtain data will require a high input of staff time and energy alongside costs of internet usage and other miscellaneous charges. This might involve setting up and maintaining a delivery team comprising of librarians to constantly look up information and deliver as soon as possible for the smooth running of the project. The study will involve continuous consultation with teaching staff to provide guidance on the direction of the project.

Reference list

Bartlett, W.A., Braga, F., Carobene, A., Coşkun, A., Prusa, R., Fernandez-Calle, P., Røraas, T., Jonker, N., Sandberg, S. and Biological Variation Working Group, 2015. A checklist for critical appraisal of studies of biological variation. Clinical Chemistry and Laboratory Medicine (CCLM), 53(6), pp.879-885.

Codari, M., Melazzini, L., Morozov, S.P. et al., 2019. Impact of artificial intelligence on radiology: a EuroAIM survey among members of the European Society of Radiology. Insights Imaging 10, 105.

Geis, J.R., Brady, A.P., Wu, C.C., Spencer, J., Ranschaert, E., Jaremko, J.L., Langer, S.G., Kitts, A.B., Birch, J., Shields, W.F., van Genderen, R.H., Kotter, E., Gichoya, J.W., Cook, T.S., Morgan, M.B., Tang, A., Safdar, N.M. and Kohli, M., 2019. Ethics of artificial intelligence in radiology: summary of the joint European and North American multisociety statement. Can Assoc Radiol J: 70(4); pp.329–334.

Gough, D., Thomas, J. and Oliver, S., 2012. Clarifying differences between review designs and methods. Systematic reviews, 1(1), p.28.

Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L and Aerts, H., 2017. Artificial intelligence in radiology, Nat Rev Cancer, 18(8): 500–510. doi:10.1038/s41568-018-0016-5

Jha, S., Topol, E.J., 2016. Adapting to artificial intelligence: radiologists and pathologists as information specialists.

King, B.F., 2018. Artificial intelligence and radiology: what will the future hold? J Am Coll Radiol, 15:501–503

Methley, A.M., Campbell, S., Chew-Graham, C., McNally, R. and Cheraghi-Sohi, S., 2014. PICO, PICOS and SPIDER: a comparison study of specificity and sensitivity in three search tools for qualitative systematic reviews. BMC health services research, 14(1), p.579.

Nabile, M., Safdara, C., John, D., Banjab, F., Carolyn, C., Meltzera., 2020. Ethical considerations in artificial intelligence. European Journal of Radiology, 122: 108768

Nadelson, S. and Nadelson, L.S., 2014. Evidence‐based practice article reviews using CASP tools: a method for teaching EBP. Worldviews on Evidence‐Based Nursing, 11(5), pp.344-346.

Neri, E., Coppola, F., Miele, V., Bibbolino, C., Grassi, R., 2020. Artificial intelligence: Who is responsible for the diagnosis?’ La radiologia medica,

Park, S.H., Han, K., 2018. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology, 286 (3): 800–809.

Pesapane, F., Codari, M and Sardanelli, F., 2018. Artificial intelligence in medical imaging: threat or opportunity? Radiologists again at the forefront of innovation in medicine. European Radiology Experimental, 2:35:

Shiraishi, J., Li, Q., Appelbaum, D and Doi, K., 2011. Computer-Aided Diagnosis and Artificial Intelligence in Clinical Imaging. Semin Nucl Med, 41:449-462

Siddaway, A., 2014. What is a systematic literature review and how do I do one? University of Stirling, 1, pp.1-13.

Speckman, R.A. and Friedly, J.L., 2019. Asking Structured, Answerable Clinical Questions Using the Population, Intervention/Comparator, Outcome (PICO) Framework. Pm&r, 11(5), pp.548-553.

Topol, E.J., 2019. High-performance medicine: the convergence of human and artificial intelligence. Nature medicine, 25(1), pp.44-56.