The use of artificial intelligence (AI) to support medical imaging analysis has been growing exponentially [1,2]. AI tools are increasingly used for a variety of tasks, ranging from automated image acquisition, sophisticated image reconstruction, classification, and diagnosis [3]. Initially, AI tools were primarily used for tasks like image segmentation or abnormality detection, whereas recent applications have advanced to deep learning–driven complex medical image classification based on large datasets, with minimal human involvement [4]. For example, AI applications have shown promising results and capabilities in the early detection of lung [5], skin [6], and breast cancers [7] from various image types such as computed tomography scans, dermoscopic (magnified) skin images, and mammography scans. In experimental studies, AI’s performance has been comparable or surpassed human readers, including experienced clinicians [8-10]. In the future, AI is expected to contribute widely to the imaging process, including image analysis and interpretation, and reduce burden on health care professionals, ultimately contributing to improved patient outcomes [1,11,12].

Although the potential benefits of AI in medical image analysis are widely recognized and its implementation is proceeding rapidly, its application in health care remains largely confined to experimental and research settings, with limited integration into clinical practice in some medical imaging applications [13]. Lack of confidence and skepticism among health care providers and patients toward novel health care technologies have been reported as contributing factors [14,15]. As technology improves, understanding if perceptions and views change is essential to develop adequate strategies to support the implementation and acceptance of AI [16].

Most existing studies and reviews to date have addressed health care professional [17] or multi-stakeholder [18] perceptions and the factors influencing their acceptance of AI in medical image analysis. Research has explored patient perceptions of AI implementation across various medical imaging domains, identifying factors such as trust, human interaction, transparency, accuracy, and individual-perceived benefits and risks associated with AI as key influences [19]. However, despite this progress, there remains a gap in understanding the full range of factors and how they interact to shape patients' perceptions of AI. Recognizing the multitude of factors and how they interact can enable implementation scientists to suggest strategies to make the AI more acceptable to patients. Furthermore, this can guide the development of information leaflets and consent forms as well as implementation strategies that are patient-centered and allow informed consent of AI use across future use cases.

The aim of this review was therefore to identify the factors that impact patients’ views of AI use in medical imaging applications and to understand how frequently they are reported, how these variables interact, or how contextual elements such as patient demographics, health issues, or the role of AI in the health care process might shape perceptions. The literature was searched with the aim to (1) identify the factors influencing patients’ perceptions of AI in medical image analysis and (2) understand how these factors influence patients' acceptance of AI.

Patients’ acceptance of AI in medical image analysis was observed to be influenced by several interacting factors. How the AI was applied in the health care setting and human involvement (ie, human-in-the-loop) was most consistently reported to shape acceptance. Hence, high acceptance was reported by studies when AI was positioned as a supportive tool assisting clinicians rather than as functioning autonomously. Other common influential factors included AI accuracy and performance, transparency that AI was used, trust in AI, clarity on who is accountable, and ethical concerns such as privacy and fairness. Sociodemographic characteristics, clinical and health experiences, and operational factors such as waiting time and cost shaped acceptance in more variable and context-dependent ways.

The level of human involvement in how AI is used had the most profound impact on patients' acceptance across studies. These findings are consistent with existing ethical guidelines and empirical studies emphasizing the importance of human oversight and accountability in AI deployment [81-83]. For example, the High-Level Expert Group on Artificial Intelligence listed “human agency and oversight” and “accountability” as a key requirement of a trustworthy AI. Furthermore, Australia’s AI ethics principles state that “People responsible for the different phases of the AI system lifecycle should be identifiable and accountable for the outcomes of the AI systems, and human oversight of AI systems should be enabled.” Furthermore, prior work has shown that acceptance of AI-supported services increases when users are informed that humans remain in control, while fear of human replacement can undermine acceptance [84-88]. This pattern is reflected in many emerging AI applications in medical imaging, where systems are typically designed to support clinician decision-making rather than replace it, such as AI tools developed for ultrasound-based disease classification and magnetic resonance imaging lesion detection [89,90].

AI-based image analysis systems are increasingly capable of high diagnostic performance, reflecting the ongoing emphasis among the emerging AI interventions to enhance diagnostic accuracy [91-93]. However, as identified in this review, although the accuracy of AI was frequently prioritized by patients, it did not override the preference for human involvement in the use of AI. Similarly, the emerging AI systems are typically designed to support clinician decision-making rather than operate autonomously, reflecting that these models account for human-in-the-loop requirement. Furthermore, similar to previous research, this review found that transparency and explainability support acceptance primarily by boosting patients’ trust rather than through inducing a detailed technical understanding of AI among them [94-96].

Studies reported that factors such as privacy and fairness influenced how patients interpret the benefits and risks of AI, thereby shaping their trust and accordingly acted as either facilitators or barriers of acceptance. Concerns related to privacy, fairness, and bias among patients mirror growing evidence that these concerns can undermine trust in AI-enabled health care [88]. Age, gender, and education of patients were the most commonly studied demographic factors; however, correlations with patients' optimism toward AI were not consistent. The inconsistent influence observed in this review aligns with literature on technology acceptance, which also found that demographic characteristics do not directly influence acceptance. Studies on technology acceptance more broadly not specific to AI found that demographic characteristics such as education shape an individual’s technological familiarity, perceived benefits, and expectations of the intervention, which ultimately influences their acceptance [97]. Despite this, acceptance can be high or low within each demographic subgroup and can change over time, as more information about a new technology becomes available and therefore needs to be assessed in future studies.

Patients’ individual characteristics like their medical history, experience with clinical contexts, and personal psychological factors were noted to act as context-setting factors, influencing expectations and sensitivities around AI. Operational factors such as waiting times and costs were identified as trade-offs, with patients potentially willing to trade-off some of their health care professionals’ involvement and accept autonomous use of AI in their screening or diagnosis for shorter wait times and reduced costs, or vice versa.

This study has several limitations. First, the conceptual model is grounded in data extracted from the included studies and informed by the authors’ interpretation; however, given the complexity of the relationships identified, not all potential interactions could be exhaustively represented. Second, the heterogeneous nature of the included studies and the use of varying survey tools limited direct comparison and quantitative synthesis. Third, certain factors such as cultural influence, socioeconomic status, and race or ethnicity were underreported across studies. This may limit the depth of conclusions that can be drawn in these areas and bias the findings toward populations and health care settings most commonly represented in the literature. This, therefore, may limit the generalizability of conclusions to more diverse contexts. Fourth, variations in terminology and conceptualization of “acceptance” across studies may affect comparability. Fifth, while the search terms and search strategy was developed with adequate guidance to be as exhaustive as possible, there is potential that some relevant publications may have not made into the search results due to variations in the keywords or controlled vocabulary used. Finally, our review focused on peer-reviewed sources reporting original research, including letters or short reports with sufficient methodological detail. Therefore, the findings may emphasize concerns that are more likely to be published, such as privacy, fairness, or specific medical contexts, and may underrepresent perspectives reported in conference proceedings, regional reports, or grey literature. This limitation should be considered when interpreting the results and designing future studies.

Overall, this review suggests that successful implementation of AI in medical image analysis should prioritize human-in-the-loop designs. Policymakers and health care organizations should focus on building governance frameworks that strengthen trust, address privacy and fairness concerns, and ensure that responsibility for AI-assisted decisions remains clearly assigned to human professionals. Furthermore, considering the complex relationship among these influencing factors, with one impacting the other in most instances, patients’ choice of accepting AI cannot be attributed to one or a few factors. Instead, promoting acceptance will require a holistic approach where all influencing factors are considered on a case-by-case basis. 

Having a standard set of constructs and variables to assess patients’ acceptance may be beneficial and enable the development of a structured approach to enhance patient acceptance of AI. The influencing factors identified in this review can form the basis for the development of standard measurement constructs to understand patients' acceptability of AI in medical image analysis for screening and diagnosis. These constructs can be further explored and substantiated by discussions and interviews with patients.

It is important to note that the relationships represented in Figure 3 should not be interpreted as causal pathways. Most included studies were cross-sectional surveys that assessed reported preferences or perceptions, often using hypothetical scenarios. As such, the model represents an interpretive synthesis of reported associations and patterns in participants’ reasoning rather than empirically tested mechanisms. The framework is therefore intended to guide future research that could prospectively examine these relationships.