Radiologists in the loop: the roles of radiologists in the development of AI applications

Abstract

Objectives To examine the various roles of radiologists in different steps of developing artificial intelligence (AI) applications. Materials and methods Through the case study of eight companies active in developing AI applications for radiology, in different regions (Europe, Asia, and North America), we conducted 17 semi-structured interviews and collected data from documents. Based on systematic thematic analysis, we identified various roles of radiologists. We describe how each role happens across the companies and what factors impact how and when these roles emerge. Results We identified 9 roles that radiologists play in different steps of developing AI applications: (1) problem finder (in 4 companies); (2) problem shaper (in 3 companies); (3) problem dominator (in 1 company); (4) data researcher (in 2 companies); (5) data labeler (in 3 companies); (6) data quality controller (in 2 companies); (7) algorithm shaper (in 3 companies); (8) algorithm tester (in 6 companies); and (9) AI researcher (in 1 company). Conclusions Radiologists can play a wide range of roles in the development of AI applications. How actively they are engaged and the way they are interacting with the development teams significantly vary across the cases. Radiologists need to become proactive in engaging in the development process and embrace new roles. Key Points • Radiologists can play a wide range of roles during the development of AI applications. • Both radiologists and developers need to be open to new roles and ways of interacting during the development process. • The availability of resources, time, expertise, and trust are key factors that impact how actively radiologists play roles in the development process.

‘If You’re Going to Trust the Machine, Then That Trust Has Got to Be Based on Something’:

“…Oakden-Rayner's critique contains important epistemological questions that deserve consideration: questions about how comparisons can be made (especially between algorithms and human experts), and how data is labelled (who labels the data, who inspects the data and whether experts with relevant clinical experience are considered). Labels and codes or criteria for comparing performances come to matter greatly when it comes to validation because they are based on the so-called 'ground truth' of features that the algorithm has learned in the training data -the labels, annotations, or codes in this instance constitute the ground truth or ground for comparison (e.g., Gulshan et al, 2016;Esteva et al, 2017;Oakden-Rayner, 2018;Cabitza et al, 2020;Scheek et al 2021).…”

“…In this article, we consider the central question of trust in Artificial Intelligence (AI) technologies for medical diagnosis. As AI becomes increasingly integrated into existing workflows and implemented to support diagnosis and treatment, clinical experts will find it difficult to understand how AI algorithms have been validated: this is where the problem of trust arises (Scheek et al, 2021). For many clinical and technical experts (such as computer and data scientists), trust is a matter of explainability and transparency of the algorithm, or the justification of the outputs of an algorith-mic model (Tonekaboni et al, 2019;Barda, 2019;Cutillo et al, 2020).…”

(De)troubling transparency: artificial intelligence (AI) for clinical applications

“…Querying data sets: 'That's where we get the ground truth label from' AI algorithms rely on data sets for training and testing the algorithm's capacity to learn. This means that the querying of data sets and process of checking the quality of datasets by clinical experts are the crucial first steps of any AI development (Laï, Brian, and Mamzer 2020;Oakden-Rayner 2017Scheek, Rezazade Mehrizi, and Ranschaert 2021;Sendak et al 2020). In our study, the querying of clinical data sets was a major recurring theme throughout our conversations.…”

“…Although AI is talked about as an already complete tool, ready for use in some domains, the development process of AI (like other technologies) plays a crucial role in laying the ground for the acceptance of the AI application (Elish 2018;Elish and Watkins 2020). In fact, the many claims regarding the positive potential of AI are counterbalanced by worries about its failings and implications, such as issues relevant to trust or mistrust (Asan, Bayrak, and Choudhury 2020;Jacobs et al 2021;Lee and Rich 2021), accountability or responsibility (Elish 2018;Lysaght et al 2019;Sendak et al 2020;Sullivan and Schweikart 2019), bias (Challen et al 2019;Cirillo et al 2020;Gianfrancesco et al 2018;Obermeyer et al 2019;Tupasela and Di Nucci 2020), healthcare data set quality (Oakden-Rayner 2017Laï, Brian, and Mamzer 2020;Scheek, Rezazade Mehrizi, and Ranschaert 2021), deskilling (Cabitza, Rasoini, and Gensini 2017;Floridi et al 2018;Laï, Brian, and Mamzer 2020); job displacement (Recht and Bryan 2017;Strohm 2019); data privacy and security (Ipsos MORI 2017;Redmore 2019). Many of these issues involve a need for transparency or the lack thereof (Tonekaboni et al 2019;Shortliffe and Sepúlveda 2018;Grote and Berens 2019;Harwich and Laycock 2018;Montani and Striani 2019).…”

Active Learning Performance in Labeling Radiology Images Is 90% Effective

“…Providing a single image with segmentation labeling can easily require 2-15 min in simple cases and longer for more complex cases, per image and per person (8)(9)(10)(11). To avoid human bias, each image is typically labeled by several human labelers with three labelers being a typical number (12,13). For radiological use cases, the labeler must be a trained radiologist, who makes this process costly and temporarily prevents the radiologist from working with patients.…”