TLDR: This research paper explores the promises, risks, and challenges of text-to-CT Scan Generative AI (GenAI) in radiology from a human-centered perspective. By involving medical students, trainees, and radiologists, the study identifies potential applications in education, training, and clinical practice, such as supplementing learning materials and aiding in diagnosis. It also uncovers critical technical challenges like image resolution and hallucinations, and domain-specific risks like confirmation bias and the need for exposure to real patient data, emphasizing the importance of stakeholder involvement in responsible AI development.
As artificial intelligence continues its rapid advancement, particularly in the realm of generative models, its integration into specialized fields like healthcare is sparking both excitement and caution. While text-to-image (T2I) generative AI models are making significant strides in creating complex medical imagery from simple text descriptions, a crucial aspect has often been overlooked: the practical benefits and usability for medical professionals themselves.
A recent research paper, titled “A Human-Centered Approach to Identifying Promises, Risks, & Challenges of Text-to-Image Generative AI in Radiology,” delves into this very gap. Authored by Katelyn Morrison, Arpit Mathur, Aidan Bradshaw, Tom Wartmann, Steven Lundi, Afrooz Zandifar, Weichang Dai, Kayhan Batmanghelich, Motahhare Eslami, and Adam Perer, this study adopts a human-centered approach to understand how medical stakeholders perceive and interact with text-to-image generative AI in radiology.
Understanding the Human Perspective
The researchers engaged with a diverse group of medical students, radiology trainees, and experienced radiologists. Instead of just focusing on technical performance, the study aimed to uncover their perspectives on the promises, risks, and challenges of a novel text-to-CT Scan GenAI model called MedSyn. This model, developed by Xu et al. (2024), generates 3D CT scans from text prompts and was integrated into a familiar open-source medical imaging viewer for the study participants to interact with.
Promises in Medical Education and Training
Participants quickly identified several promising applications for this technology in medical education and training:
-
Supplementing Learning Materials: The AI could generate specific examples for lectures, allowing students to interact with and compare normal and abnormal anatomies, which is often difficult with static images or limited real cases.
-
Expediting Learning: Trainees could generate variations of diseases, explore differential diagnoses, and visualize rare conditions that they might not encounter frequently in real practice. This exposure could significantly accelerate their learning curve.
-
Empowering Trainees in Emergencies: For those on-call or in high-pressure emergency situations, the tool could serve as a quick reference to generate images of critical conditions, helping them make faster, more confident decisions.
Applications in Clinical Practice
Beyond education, the study also highlighted potential uses in daily clinical practice:
-
Improving Report Impressions: Radiologists could generate reference images based on their interpretations to visualize hypotheses, refine diagnoses, and add more specific details to their reports. The ability to compare generated images side-by-side with real patient scans proved particularly insightful for some participants.
-
Planning and Communication: The generated CT scans could be valuable for surgical planning or for explaining conditions to patients, offering a visual aid to communicate complex medical information more effectively. They could also serve as additional evidence when discussing diagnoses with other medical professionals.
-
Visual Memory Support: For experienced radiologists, the tool could help recall visual features of pathologies not encountered recently, acting as a quick alternative to traditional online search engines.
Identified Risks and Challenges
Despite the excitement, the human-centered approach also brought to light critical technical and domain-specific challenges:
-
Technical Challenges: Participants noted issues with the resolution quality of generated CT scans, the presence of “hallucinations” (incorrect or injected pathologies not requested in the prompt), the model’s inability to produce contrast-enhanced CT scans, and the significant generation time (3-4 minutes per scan).
-
Domain-Specific Risks: A major concern was confirmation bias, where users might unintentionally see what they expect to see in a generated image, potentially leading to learning incorrect information. Another significant risk was the over-exposure to synthetic images, emphasizing the ongoing need for medical professionals to continuously interact with real patient CT scans to develop accurate diagnostic skills.
Also Read:
- GeMix: Enhancing Medical Image Augmentation with Learned Generative Mixing
- Advancing Breast Cancer Screening with AI: A New Method for Generating Paired Medical Images
Implications for Responsible AI Development
The study underscores that developing advanced AI models for high-stakes environments like medicine requires a deep understanding of user needs and potential safety risks. The findings highlight the importance of human-centered evaluations to uncover challenges that quantitative metrics alone might miss, such as the subjective nature of CT scan impressions or the specific types of biases that concern radiologists.
Future work should focus on designing AI applications that enable variability in generated images, explicitly identify synthetic content to prevent misuse, and develop visual explanations to safeguard against hallucinations and confirmation bias. This research serves as a crucial step towards ensuring that medical text-to-image GenAI is developed responsibly, aligning with the complex workflows and critical needs of healthcare professionals. You can read the full research paper here.
