Impact of artificial intelligence on radiology: a EuroAIM survey among members of the European Society of Radiology

Radiology generates a huge amount of digital data as obtained images are included into patients’ clinical history for diagnosis, treatment planning, screening, follow up, or prognosis. Besides, the increasing use of computers and data has led to the successful utilisation of artificial intelligence (AI) to carry out several tasks for more accurate and up-to-date results. The European Society of Radiology (ESR) put together a survey aimed at determining the radiologists’ position towards these new technological innovations which could strongly impact their specialty.

About 2.8% of the 24,000 contacted ESR members answered the survey entirely. According to respondents, breast, oncologic, thoracic, and neuro imaging are the most likely to be strongly impacted by AI and new technological innovations, along with forms of imaging such as mammography, computed tomography (CT), and magnetic resonance imaging (MRI). More than half of respondents anticipate patients not to accept AI-only based reports while 12% expect patients to accept these reports and 33% stated it’s too early to answer this question. Meanwhile, it is still unclear if the responsibility of the AI systems outcomes will be borne by the radiologist alone or if it will fall under a shared responsibility scenario.

Radiologists’ job opportunities and workloads are expected to increase or decrease. The respondents believe that, should AI allow to save working time, the saved time should be used to provide for stronger interactions and increased communication with other clinicians (64.7% of respondents) and patients (47.7% of respondents). Of all radiologists who answered the survey, 41% believe the relationship between them and the patient will become more interactive while 21% claim it will become more impersonal. The remaining respondents believing the relationship will remain unchanged.

ESR respondents unanimously agreed that radiologists must play a leading role in developing and validating AI applications to medical imaging. It will require significant involvement of radiologists and use of their expertise to ensure the quality of data and effective transformation of development solution research into clinical practice. This, with the aim of improving the outcome for patients and the trust of patients in new developments. Moreover, accountability and ethical issues surrounding AI-systems would constitute a significant challenge that would require regulations at EU and international level. The survey also highlights the radiologists’ will to be educated on advantages and limitations, the clinical use, and technical methods of AI applications.

Among medical subspecialties, radiology is one of the most prolific generators of digital data. Each radiology department routinely generates a large and heterogeneous amount of data. Images daily acquired with different modalities, such as radiography, angiography, ultrasound (US), CT, MRI, or nuclear imaging, are integrated to patients’ clinical history to extract information with the aim of screening, diagnosis, treatment planning, and prognosis [1].

In this digitised environment, AI found its fertile ground for flourishing. Indeed, the recent achievements obtained thanks to the application of machine learning for medical image analysis have shaken the radiological world [2, 3]. Machine learning proved its applicability to different imaging modalities and radiological subspecialties. In particular, deep learning has emerged as a promising technique for processing medical imaging data, being employed for several tasks like image classification, segmentation, registration, and abnormality detection [4].

This data-driven revolution has the concrete possibility to drastically impact on radiology. The debate about how abrupt this impact will be is still open and controversial. Indeed, several opinion articles have been published reporting experts’ forecasts about the possible scenarios that radiologists will face in the next future. On the one hand, there are the optimists, who see in the AI an opportunity to enhance radiologist’s role in the healthcare system; on the other hand, there are the pessimists, who predict a relatively fast replacement of radiologists by AI systems [2, 5, 6].

Even though some of these scenarios may be considered as extreme, they all have a common characteristic: the involvement of radiologists as one of the designated users of AI-based systems. The controversial part is related to their active or passive role in this healthcare revolution. It is not easy to predict reactions and attitudes of radiologists to this technological innovation but we can investigate their current feeling about what the future will hold.

In this article, we report the results of an online survey aimed at investigating the feelings of members of the ESR about AI impact on their practice.

We conducted an online survey entitled “Your expectations about AI in radiology”. The survey was composed of two subparts. The first part consisted of 7 questions related to respondent age, sex, radiology subspecialty, most frequently practiced techniques and working status, type of institution, and country (Table 1). No personal identifying data were collected.

The second part consisted of 15 multiple choice questions about user feelings/forecasts in respect to the advent of AI applications in radiological practice in the next 5–10 years. The detailed list of questions and possible answers is reported in Table 2. The survey was designed by the European Network for the Assessment of Imaging in Medicine (EuroAIM), a joint initiative of the European Institute for Biomedical Imaging Research (EIBIR), under the umbrella of the ESR Board of Directors.

The online survey was designed using SurveyMonkey platform (SurveyMonkey, San Mateo, CA, USA) and distributed by an email containing a non-serialised URL link. The email was sent to all ESR members. The survey was opened on November 6, 2018. The first email was then followed by two reminders on November 20 and 26, 2018. The survey was closed on December 8, 2018.

Results of ESR members responses to this survey were automatically recorded and processed using Excel (Microsoft, RedMond, WA, USA) and Matlab r2018b (MathWorks, Natick, MA, USA). Since all the questions deal with categorical variables, descriptive statistics was reported by means of frequencies and percentages. Moreover, correlation between ESR members and survey responders per country was estimated using the Spearman’s correlation coefficient.

In one month, 675 ESR members, most of whom are radiologists aged from 30 to 60 years, working in academic or non-academic hospitals, completed the proposed survey. The total number of responders is small if compared to the total number of ESR members. However, the number of responders per country correlates to the number of ESR members per country. The prevalence of colleagues based in academic/teaching hospitals (over 50%) should be related to the higher sensitivity to AI innovation in these centres.

Radiologists who answered the survey practiced different subspecialties. Among possible options, none prevailed. General and abdominal imaging were the most practiced subspecialties by responders, each of the two accounting for 29% of responders (see Fig. 3). In spite of the subdivision into different radiological areas, we observed a quite large agreement among them regarding which subspecialties will be more influenced by AI applications. Indeed, more than half of responders stated that breast (61.6%) and oncologic imaging (53.6%) will be most impacted by the AI revolution, followed by thoracic imaging (42.2%), neuroradiology (32.7%), and nuclear imaging (21.6%). This is reasonable, because consolidated screening programs provide, since the beginning of AI era in the healthcare, a large amount of digital data. This condition makes breast imaging the first candidate for the application of machine learning algorithms [5]. It does not surprise that the highest rated imaging subspecialties are those that frequently involve tumour detection and characterisation. Indeed, this task represents a classification problem, which is prone to be solved using machine learning algorithms [7]. Notably, radiomics represents the most promising approach for characterisation of solid cancers, which are spatially and temporally heterogeneous [8, 9]. Machine learning algorithms take advantage of the heterogeneity of imaging data used for cancer patient diagnosis and treatment. Imaging protocols traditionally used for cancer patient management include both morphological and functional imaging, which can be successfully processed using machine learning algorithms. These algorithms take advantage from the volume and heterogeneity of information contained therein to detect specific patterns starting from raw data.

Most of responders practice CT and MRI, accounting respectively for 93.6% and 84.9% of responders. Other frequently practiced modalities are US (72.0%) and radiography (69.9%). As expected, when asked to foresee which imaging modalities will be the target field of application of AI systems, responders suggested CT (79.7%) followed by MRI and mammography (56.0 % each). We should note that CT and MRI are only relatively standardised modalities while mammography (at least in the screening setting) is highly standardised. Conversely, US remains poorly standardised in every field, adding a confounding variability to the data that machine learning should model (except for automated breast US, still in its investigational phase). In fact, US was not selected as a probable field of AI application in the next 5–10 years: only 3.7% of responders foresee AI systems implementation on US practice.

It does not surprise that many responders suggested that the detection of disease in asymptomatic subjects (60.1%), staging/restaging in oncology (46.5%), quantitative imaging biomarkers (37.9%), and image postprocessing (35.9%) should be regarded as target application of AI systems. In support of this claim, it should be noted that machine learning algorithms showed valuable alternative to conventional image processing approaches in several tasks, such as image segmentation, registration, classification, and enhancement, making them the most promising approach in computer vision and medical image processing [4, 10].

Radiologists foresee an impact of AI systems on their job opportunities (55.6% of responders) and reporting workload (74.7%). However, whether this impact is felt as an increase or a decrease is unclear: responders forecast both scenarios with similar probabilities (58.1% versus 41.9% for increase/decrease in job opportunities¸ 50.8% versus 49.2% for increase/decrease in reporting workload). This fact points out radiologists’ uncertainty about their professional future in the AI scenario. At present, there is no agreement on how AI will affect the labour market or the workload of a radiology department due to the currently limited application of AI systems in daily clinical practice. In fact, only 20.4% of survey responders are currently using AI-based application in their practice. Curtis P. Langlotz [11] recently documented how the predictions of a fast disappearance of radiologists from the healthcare systems (“the mirage of job displacement”) are very far from real clinical practice for now and the predictable future for many reasons, including the “long tail” of the disease distribution, accounting for rare diseases. Taking in mind the example of the autopilot for human pilots, he repeated the simple true affirmation: Radiologists who use AI will replace radiologists who don’t.

The percentage of responders believing that AI will make their duties more clinical (53.9%) or unchanged (14.2%) is higher than those who foresee a more technical profile (27.7%). In addition, the percentage of responders believing that AI will make radiologists more subspecialty-focused (41.9%) or not different from now (40.1%) is higher than those who foresee a less subspecialty focus (17.9%). Thus, AI is mostly perceived as an innovation favouring a higher clinical profile and subspecialty dedication of radiologists, which are key factor for being visible to colleagues and patients, for instance playing a pivotal role in multidisciplinary tumour boards. In fact, the opposite AI-driven scenario would be radiological reporting as a low-value commodity.

As a further demonstration of a positive perception of AI by the responders, they consider other clinicians (64.7% of responders) and/or patients (47.7% of responders) the main target for using the time potentially saved by using AI, thus increasing the quantity and quality of clinical communication with them. In fact, 40.7% of responders think that AI will make the radiologist-patient relation more interactive and only 20.7% think that AI will make the radiologist-patient relation more impersonal (38.4% do not foresee any change). Overall, there is a prevalence in favour of AI as a factor making the radiologist’s role more clinical and patient-oriented.

The debate about who will take the responsibilities of AI outcome is controversial [12]. Of note, this is a central issue because it may drastically affect the role of radiology and radiologists in the healthcare systems. The results of the survey are somehow unexpected. On the one side, 55.4% of responders believe that patients will not accept a report made by an AI-application alone (without supervision and approval by a physician) and only 11.7% claim the opposite (32.9% believe it is now too early to estimate patients’ reaction). However, only 41.1% of responders identify the radiologists as the only professionals who will take the responsibility of AI outcome, 41.1% would favour a “shared responsibility”, with a minority of responders proposing AI developers (10.2%), or insurance companies (3.6%). We interpret this data as a word of cautiousness. We are at the beginning of this revolution. To take the responsibility of AI outcome implies many things up to now not clearly defined, first of all the potential for modulating the AI outcome in relation to the specific patient, to correct or declare a different opinion, to discuss with other clinicians a given outcome (think about the imaging-based declaration of disease progression during an anticancer treatment). These results clarify that accountability of AI-systems output still represents an open challenge that will require regulations at the European level and across the countries.

An important result of this survey is that 100% of responders believe that radiologists must be involved in AI-system development and validation. Radiologists are eager to be protagonists of this revolution and to manage all the steps from the development to the application of AI systems. To reach this scope, they suggest that training programs should be adopted to teach trainees and clinical radiologists advantages, limitations, and clinical use of AI-based systems. Unfortunately, only 30% of them considers the hypothesis to play a key role in the creation of labelled dataset. As a matter of fact, the accuracy of data-driven AI systems strictly depends on data quality [13]. Without radiology expertise, to expect an accurate outcome and consequently a successful translation of AI systems from research setting to clinical practice is only a wishful thinking. The radiologist’s role in this technological revolution is fundamental. Data-driven approaches require someone that ensures the quality of data, especially when supervision is needed during training of machine learning models. In addition, radiologists’ expertise in diagnosis, treatment planning and prognosis, and related issues, is the cornerstone of any successful translation of developed solution from research to clinical practice. Radiologists should flank engineers, computer scientists, and other people involved in the development of a decision support system, with emphasis on task definition and high-quality image labelling.

This report has two main limitations, already mentioned at the beginning of Discussion. First, the limited number of responders compared to the ESR members and the number of members who were reached by the emailing system. Then, half of responders work at university or teaching hospitals, so the answers reflect the opinions and feelings of advanced realities, not of the general world of radiologists. However, the analysed answers give an idea of that part of the radiological world that has a guiding role and also teach radiology to residents who are the future of radiology.

In conclusion, the responders to this survey showed on average a favourable attitude towards the adoption of AI-systems. A positive scenario of a more clinical and patient-centred radiological profile in the AI environment is foreseen. However, radiologists still do not know how fast and how disruptive the implementation of machine learning systems in radiology will be. This explains some cautious answers on the role of radiologists in responsibility for AI outcome and in the development of AI systems, considering that many ethical and legal issues related to the use of these systems are still unsolved. The future will depend on what we are doing now and what we will do in the near future, i.e., on our ability to exploit the many opportunities that AI will offer to radiology and radiologists [14]. As Alan Key said: The best way to predict the future is to invent it [15].