Generative Artificial Intelligence in Nuclear Medicine Education

Cautions

There remains debate across the education sector about the benefits of text-to-text genAI (e.g., research, writing, and problem solving) versus misuse (scientific fraud, cheating, academic integrity). In parallel, the rapid assimilation of text-to-text genAI in clinical, research, and academic environments has occurred ahead of the development of regulatory and ethical frameworks. It should be a priority to develop appropriate use and ethical guidelines for application of genAI that are valid across the nuclear medicine and educational domains. This is reflected in a recent systematic review that revealed that perceptions of the use of ChatGPT centered on perceived benefits to academic and scientific writing and to scientific research, countered by concerns about ethical issues, risks of use, and inaccuracies (6). The negative perception of genAI in education and scientific writing is perhaps skewed by the term artificial intelligence, which implies something false when, in reality, the role of AI and genAI in these domains is augmentation of learning and capability, or engineered learning. When genAI enhances learning and capability, it should be supported, but when it is used to mask a lack of learning and capability or to craft the impression of evidence of learning and capability, it should not be acceptable.

ChatGPT is helpful when shaping the structure of a piece of writing for students or a piece of scientific writing but lacks the higher-order capabilities required for synthesis and communication of complex medical or technical information. For example, ChatGPT is effective in providing a generic explanation of a nuclear medicine procedure but lacks the reasoning to navigate the nuances of medical language to accurately interpret meaning from a medical report. GenAI is rapidly improving, and recent versions of ChatGPT included higher reasoning capabilities. ChatGPT is also prone to hallucinations, confabulations, and delusions that embed errors in responses (7). Consequently, although sounding convincing, ChatGPT responses are inadequate for guiding students or educators or for analyzing and interpreting clinical or research results. Lower levels of understanding among students are likely to make it more difficult to identify hallucinations in responses. For example, the general and superficial nature of GPT-3.5 and GPT-4 training was shown to be inadequate to pass undergraduate radiography and nuclear medicine examinations that were structured to draw out understanding, but GPT-3.5 and GPT-4 performed better than students in shallow fact-based questions and multiple-choice examinations (8–10).

Text-to-text genAI algorithms have also been shown to reinforce and amplify historical and institutionalized biases that threaten sex, ethnic, cultural, and social diversity and inclusivity (11). This is an important principle for those adopting genAI use in nuclear medicine education because the inherent biases would not be acceptable from the human user in the user’s own interactions and communications and are counter to diversity and inclusivity strategies across professions and organizations. Importantly, inherent biases can form part of the hidden curriculum if used in education and can reinforce stereotypic perspectives of professions and professionals. ChatGPT, for example, reinforces professional sex stereotypes through depiction of doctors as male and nurses as female (12).

Applications

Although GPT-4o offers improvements over GPT-3.5, accessibility is an important consideration in educational use. Inequity will result among students if there is a reliance on GPT-4o and the associated paid subscription. When ChatGPT is used as a tool in education, it is important to either supply the GPT-4o response or implement activities that are equally appropriate on GPT-3.5. Perhaps the most significant barrier to more widespread adoption in nuclear medicine education is the risk of AI-enabled scientific fraud and academic misconduct. GenAI can synthesize data, medical histories, patient reports, and other deep fakes, including student assessments. Although these concerns present risk, they can also present rich learning opportunities. I use the following applications of text-to-text genAI to enhance the learning of undergraduate nuclear medicine students.

Text-to-text genAI can be used with confidence to support the development of student writing skills. The training set does not support the depth of insight and understanding expected of a university student and, therefore, does not produce written work at a pass level (8,9). This relies on appropriate alignment of the taxonomies of the learning outcomes with the criteria in the assessment rubric. What genAI is useful for is to help students develop their writing skills and writing style. Students can craft, for example, an assignment section and prompt genAI to reword for better flow and structure. The value comes from students reflecting on suggested changes and taking notice of the justification that ChatGPT, for example, provides for the edited version (e.g., “by reorganizing the information and using a more cohesive narrative, this version provides a clearer explanation”). This is particularly useful for students transitioning from high school to university studies and enriches the feedback provided by the examiner. The genAI interaction and learnings should be encouraged, with reflection, to be included in a student portfolio as the repository of evidence of student learning. Similarly, another authentic application of genAI among student writing tasks is to suggest a structure for a written piece. For example, asking for a suggested structure for a particular task could include details of the task expectations to help guide student research and direction and decrease the chance that students will omit key research and associated learning. In both cases, the work remains that of the student, with genAI augmenting performance in an authentic way. That is, genAI is used in a manner similar to the way in which it might be used in clinical practice (e.g., a practitioner asking for a report, letter, or abstract to be reworded or asking for the structure for a research report, grant application, or journal paper).

A particularly useful way to use text-to-text genAI in education is to prompt a question and provide students with the response. This could be in a class tutorial environment, group activity, or examination. The idea is to have students demonstrate their understanding of the topic by critiquing the genAI response. This approach leans on genAI’s making errors or including confabulations, hallucinations, and delusions, thereby affording students the chance to show depth of understanding. As a bonus, the task also gives students the opportunity to become acutely aware of genAI inaccuracies, which, in turn, educates students away from misuse. GenAI can be prompted to include several general or specific errors and to explain the errors to help the academic in marking them.

A valuable application of genAI is the initial development of communication and interpersonal skills, including interprofessional interactions. Although authentic interactions with patients and practitioners provide the richest learning, genAI affords the chance for students to simulate interactions. This better prepares students for clinical placement but also removes some of the preliminary teaching from the clinician, which is valuable given workforce pressures. For example, prompting genAI to play the role of a patient about to have a nuclear medicine procedure while the student has a conversation with the patient and answers patient questions is a simple simulation. The student can reflect on the patient’s responses and can refine and run the scenario multiple times. The same approach can be adopted for interprofessional learning, in which genAI plays the role of one or more characters. For example, students may be required to resolve a conflict through discussion with a patient’s referring clinician and the nuclear medicine physician. Not only can the scenario be repeated to try to improve outcomes, but hyperparameters can be adjusted to increase authenticism. That is, in the example above, a terse tone could be attributed to the nuclear medicine physician and a condescending tone to the clinician, leaving the nuclear medicine student to navigate that conversation. Similarly, patient conversations can include a tone that creates an anxious, angry, overly friendly, confused, forgetful, or child conversational avatar. The conversation can be shaped by different levels of health literacy attributed to patients. As with the written tasks, the richness of learning comes from students reflecting on their learning and including those reflections in a portfolio of evidence.

Scientific fraud using text-to-text genAI is a significant issue, but this capability can also be harnessed for enhanced student learning. Abstracts of articles shaped to provide specific teaching points can be fabricated and used to build critique and critical thinking skills among students. Entire research datasets can be generated by genAI that match a specific outcome. These can be used, ethics-exempt, by students to develop or assess analysis and interpretation skills. There are many other purported benefits of text-to-text genAI that are all subject to errors and are counterintuitive to the actual benefits proposed. ChatGPT can provide a list of its own capabilities, but most are not at the actual depth or currency of knowledge and level of understanding expected of a university student in nuclear medicine.

Limitation

An important limitation is that genAI can appear to change mood and responsiveness. For those who have used ChatGPT, for example, in a single session, multiple sequential prompts can have similar forms and structures before a very dramatic change. It almost feels as if a human has been answering questions and then a change in shift has a different person take over. Of course, the genAI responses and tone can be shaped by the user prompt, another valuable tool for student simulation. Responsiveness between sessions is also very different, and some of the simulations described above can have variable success. For example, in one session, ChatGPT creates a conversational avatar that adopts the role of 1 or 2 characters and interacts in a simulated conversation. In another session, the same prompt confounds ChatGPT and it simply provides dialogue for all characters with no input from the user. This can frustrate students, produce variable learning experiences, and undermine the value of genAI in student learning.

Cautions

GANs are used to synthesize images to enrich datasets for research and training but usually have a fairly specific application and training set. This makes them less useful for more general image production for educational purposes. The recent emergence of text-to-image genAI using diffusion models has propagated considerable interest in potential applications in education, including in nuclear medicine. The key caution for use in nuclear medicine education relates to image quality. Image quality can be considered against several criteria (5,13). The first criterion is the alignment of the image produced with the intention of the prompt (image-to-text alignment). In Figure 5, the prompt required a photorealistic image of a radiographer performing an x-ray on a patient’s arm. Ignoring the variable image quality in images produced by DALL-E 3, Firefly 2, Stable Diffusion 2.1, and Midjourney 5.2, the images show a radiographer, but could any be interpreted as performing an x-ray on a patient’s arm? Image quality has been listed as the second criterion and will depend on the prompt requirements. For example, the prompt could request anime style characters or more abstract images. In Figure 6, the prompt required photorealism. Firefly 2 relies on stock images, and if an appropriate character is contained within those stock images, it produces the most photorealistic result. Midjourney 5.2 tends to produce the most photorealistic completely synthetic images, followed by DALL-E 3, whereas Stable Diffusion 2.1 tends to be of low quality. The third criterion is visual reasoning, which is a reflection of the accuracy of object recognition (are objects in the foreground and background identifiable?), object counting if specific values are prompted, the context and scene, and spatial relationships between objects. In Figure 6, for example, Firefly 2 has a bandless watch and chest x-ray. At the junction of image quality and spatial reasoning is a fourth criterion associated with distortions, artifacts, and errors in image formation. Some of these distortions, artifacts, and errors can be obvious and dramatic or fairly subtle, but users should keep in mind that the keen observer will identify anomalies and there is little control over how that content is then represented, including potentially maliciously.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5.

Comparison of variability of quality criteria for text-to-image genAI using DALL-E 3 (top left), Firefly 2 (top right), Stable Diffusion 2.1 (bottom left), and Midjourney 5.2 (bottom right).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 6.

Five serial images from DALL-E 3 in response to prompt for images of typical nuclear medicine technologist. Diversity and inclusivity of medical imaging professions are not represented in these images.

Perhaps the most important criterion is bias. Like text-to-text genAI, text-to-image genAI is prone to biases that may cause it not to reflect the diversity and inclusivity of nuclear medicine—not a message to instill among nuclear medicine students. DALL-E 3 was prompted through multiple iterations to create images of typical radiographers and nuclear medicine technologists and produced predominantly young men with a light skin tone (Fig. 6). Sex and ethnicity biases were also revealed for DALL-E 2, Midjourney, and Stable Diffusion among surgeons and ophthalmologists (14–16). Diversity and inclusivity are important to instill among students, but if the resources that are used reflect a less positive message, these will drive inherent biases among graduates and externally portray the nuclear medicine professions erroneously, undermining industry recruitment and workforce sustainability.

Applications

Despite some of the issues with image quality and biases, there are numerous applications of text-to-image genAI in nuclear medicine education, such as visually representing anatomy or physiology; analyzing images, including medical images, to assist students in detecting and identifying normal structures, disease, or anomalies; producing authentic medical images (e.g., x-rays or CT scans) or pathologic images as teaching examples; generating promotional or marketing material to promote courses; producing patient-facing education or information posters as assessment tools for students; producing professional images for circulation in social media; and producing images that showcase inclusivity and diversity in the professions. Several of these applications have been evaluated in the scientific literature.

Creation of anatomic images by DALL-E, Stable Diffusion, and Craiyon V3 was assessed, with no images providing sufficient detail or accuracy to be considered useful (17). Similar observations were made for genAI-crafted images of pathologic conditions (17). Radiologic images produced by DALL-E 2 superficially looked realistic but lacked the detail required for educational value (18). All my efforts to integrate text-to-image genAI into student learning have been confounded by inaccurate, inadequate, and inappropriate images unsuitable for use as visual tools or for student guidance for revision and learning. Although the purported applications of DALL-E 3 are conceivable, the current capability leaves these largely as future considerations, and use in current practice is fraught with risk.

Among the purported applications, those supported by anecdotal evidence are limited. Students are visual learners, and the use of visual stimuli for simulation and scenario-based learning is beneficial (Fig. 7). The simulation laboratory itself, including where text-to-text genAI communication simulations are used, may not provide the visual context to drive fully immersive simulation. In the absence of virtual reality, specific scenarios can be generated for student simulation and, indeed, class discussion and debate. This is particularly valuable in the online learning space, where genAI images can be added to stimulate class discussion. This could include a realistic image depicting a scenario (e.g., a scene with an anxious parent of a pediatric child) or one with inaccuracies or issues students are expected to identify. Again, this activity can integrate with text-to-text genAI simulation. The same approach could be adopted to provide visual context for an ethical discussion or debate. A useful tool for genAI is the crafting of images that represent the profession and their use in social media or promotional material.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 7.

DALL-E 3 image application in nuclear medicine student education. On left is visual stimuli that could be used for class discussion and debate or to set scene for hands-on simulation practicals. On right is visual stimuli that could be used in class or online to generate discussion and debate around ethical dilemmas.

Limitation

The obvious barrier to text-to-image genAI is that the algorithms are not ready for the depth and accuracy required to represent medical-type images. This includes the algorithm biases and the image aberrations. Nonetheless, these limitations and criticisms represent the status at the time of writing. The space is rapidly evolving, and the horizon shows great promise for higher-quality images that could reengineer the way nuclear medicine education is delivered. This is potentially transformational from a sustainability perspective, with the potential—on the horizon—to generate complete teaching case sets and simulation activities without the traditional costs of simulation systems. With improved algorithms will come increased capabilities, but if only accessible through paid subscriptions, then there is inequitable access for students unless an institutional license is purchased. For DALL-E 3, individual students would need a subscription to GPT-4 or GPT-4o (current at the time of writing) or access to DALL-E 3 through an institutional Microsoft Copilot subscription that is extended to students (rather than just staff). Similarly, Midjourney requires subscription access. The more freely available applications, such as Stable Diffusion via HuggingFace, have limited quality and usefulness. Firefly 2 could be institutionally available through an Adobe subscription, but the full gamut of tools may not be available to students or staff. There remain several useful applications by which university academics craft genAI images for visual prompts in learning, but student inequity must be considered when student-driven genAI tasks are set.

A complicating or confounding factor is the censorship on medical images generally applied by algorithms. This is designed to prevent inappropriate or malicious use of potentially sensitive images or images that could cause confusion, misunderstanding, or community concern. For example, Midjourney has algorithm bans on medical terms, limited medical stock images, and variable quality of medical image outputs (19). DALL-E 3 and Firefly 2 similarly may respond to a text prompt by indicating that the content of the text prompt breaches guidelines. For example, a prompt to generate an image of a patient with cancer in DALL-E 3 could be met with an error associated with inappropriate content or may produce an image of a patient who appears to have cancer. Similarly to ChatGPT text-to-text genAI, the DALL-E 3 algorithm seems to have moods and may be more cooperative with the same prompt at different times. There are 2 ways to circumvent, in some cases, errors associated with medical content. The first is to simply add an extension at the end of the prompt that says “for educational purposes.” The second is perhaps more fitting in nuclear medicine education. The original prompt can provide more specific detail of what is expected from the image. For example, the algorithm is more likely to comply with a prompt to generate an image of a cancer patient if the prompt also adds the context of portraying strength, courage, resilience, hope, and a supportive environment and will almost certainly produce a more representative image. Indeed, astute prompt engineering for text-to-image genAI is the most powerful weapon in crafting representative images free from biases and context misalignment. Yet this takes patience and multiple iterations of both the prompt and the image, a requirement that is counter to the expectations of immediate gratification associated with genAI use.