Artificial Intelligence-Generated Facial Images for Medical Education

Abstract

We evaluated the use of text-to-image models (Microsoft’s Bing Image creator (powered by DALL·E) and Shutterstock’s AI image generator) to generate realistic images of human faces and their associated pathology, which may be useful for medical education, given they may overcome issues of patient privacy and requirement for consent. These models have potential to augment rare medical image datasets for medical education, as well as provide greater inclusivity and representation of diverse populations.

Artificial intelligence (AI)-powered text-to-image models that generate images from textual prompts such as DALL·E-2 have received global attention for their ability to create images ranging from simple depictions to more creative and imaginative interpretations based on the given input. The ability to generate images from textual prompts has diverse applications, including content creation, visual storytelling, artistic expression, and design assistance. With rapid developments in generative AI and accompanied by increasing computing power, these models are poised to bring about quantum leaps in multiple industries, including healthcare and medical education. Specifically, AI can be utilised to improve medical education, both in the creation of educational imagery and narrative medicine [1]. The application of images for medical education through lecture presentations, clinical case and small group discussion for both basic science and clinical teaching is rooted in principles of cognitive science, learning theory, and the unique demands of medical instruction. Visual materials enhance learning by simplifying complex information, reducing cognitive load, promoting engagement, and aiding memory and recall. They play a crucial role in preparing future healthcare professionals to understand and navigate the complexities of the human body and medical practice.

Photographs of patients are essential in medical teaching [2], where they capture classical phenotypic features, enabling students to recognise diseases and providing an important gestalt that written or verbal descriptions cannot give. Historically, clinicians and medical journals have endeavoured to protect the anonymity of patients. However, this has shifted towards obtaining full consent for all images, with a consensus by the International Committee of Medical Journal Editors implementing a policy of full consent for images [3]. Moreover, medical images are subject to the Health Insurance Portability and Accountability Act of 1996 (HIPAA), a US federal law enacted to regulate the collection and use of sensitive patient health information and prevent images from being disclosed without the patient’s consent or knowledge. We recognise that while patients may be comfortable with photographs that are used exclusively for their personal medical records, the extension of its use for medical teaching or publication may not be so readily accepted [4]. More recently, the sharing of medical images by medical students by smartphone on Figure 1 (https://www.figure1.com), a mobile app and online platform designed for healthcare professionals to share and discuss medical images, raised concerns if the terms of use were adequate to protect patient privacy [5]. The generation of artificial facial photorealistic images addresses the main concern of patient privacy and the violation of patient autonomy, where improper use of patient images erodes the rights of individuals to self-determination and society’s trust in the medical profession.

We utilised 2 existing text-to-image models available in the public domain, Microsoft’s Bing Image creator (powered by DALL·E) and Shutterstock’s AI image generator, for purposes of generating photorealistic facial images for medical education. The text prompts “baby face with cleft lip surgical repair,” “bruise asian female face,” “burns affecting face,” and “Bell’s palsy” were used. Results are shown in Fig. 1: Fig. 1A shows a Hispanic baby’s face with post-surgical repair of a right upper cleft lip defect; Fig. 1B features an East Asian adolescent female pained face with a right facial bruise with furrowed eyebrows; Fig. 1C depicts a young Caucasian female with a left facial third-degree burn with frank blistering and charring of the face; lastly, Fig. 1D portrays a middle aged Caucasian male with a right facial nerve palsy as shown by the loss of the right nasolabial fold. The text-to-image models generated realistic images of human faces and their associated pathology. AI-powered text-to-image models may also help to augment rare medical image datasets (such as images of children with facial trauma from non-accidental injuries—Fig. 1B and C could represent faces of young vulnerable females subject to domestic violence), as well as provide greater inclusivity and representation of diverse populations [6, 7].

Fig. 1.

AI-powered text-to-image models

However, there are significant limitations to current text-to-image models when tasked specifically with medical image generation. Existing models are trained upon non-medical, large-scale image datasets such as ImageNet, COCO (Common Objects in Context), Open Images and a broader range of internet images, which may impede the learning of visual representations and patterns, and final generation of images specific to the medical domain. Hence, the accuracy and reliability of such generated images need to be carefully validated and verified by medical experts and educators before use. Additionally, ethical concerns related to data privacy, consent, and potential biases in the training datasets should be addressed before utilising AI-generated images in medical education. We recognise that while text-to-image models show great potential as an education tool, key issues particularly harm-by-data, given the existing algorithmic biases need addressing, where minority groups, race, gender, vulnerable social classes, and lower socioeconomic status face data-marginalisation due to underrepresentation in existing training image datasets. We envision a future where text-to-image models trained on globally sourced image datasets, tempered with good AI governance involving human agency and oversight, privacy, and data accountability, could generate facial images that encompass a wide range of ethnicities, ages, genders, and physical attributes, thus promoting diversity and cultural sensitivity in medical curricula.

Funding

BEF is supported by the National Medical Research Council Grant (NRMC/CIDA19May-0004).

Data Availability

All data generated or analysed during this study are included in this published article.

Declarations

Consent for Publication

Not applicable (as images are AI-generated).

Competing Interests

The authors declare no competing interests.