Related Experiment Video
Updated: Jul 3, 2026
05:23
A Lip Fixation Device Facilitates Labial Salivary Gland Biopsy for Sjögren's Disease
Published on: March 6, 2026
Evaluation of generative artificial intelligence in producing anatomically distinct lipedema subtypes: A diagnostic
Ilhan Celil Özbek1, Bülent Alyanak2, Burak Tayyip Dede3
1Department of Physical Medicine and Rehabilitation, University of Health Sciences Derince Training and Research Hospital, Kocaeli, Turkey.
Phlebology
|July 2, 2026
Summary
Generative AI can create realistic medical images but struggles with lipedema subtypes. The AI model accurately depicted Types I-III lipedema but failed to generate images for Types IV and V, limiting its use in medical education.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Dermatology
Background:
- Generative AI models are increasingly used for medical image creation.
- The accuracy of AI in representing distinct disease subtypes, like lipedema, is not well-established.
- Lipedema has five distinct anatomical types according to the Schmeller classification.
Purpose of the Study:
- To evaluate the diagnostic accuracy of a generative AI model in producing images of the five anatomical lipedema subtypes.
- To assess the AI's ability to differentiate between lipedema Types I-V based on anatomical representation.
Main Methods:
- A prospective audit used ChatGPT's image-generation interface to create 60 images for each of the five lipedema types (total 300 images).
- Prompts were standardized to include only the lipedema subtype label.
- Two independent clinicians classified the generated images, with disagreements resolved by a third clinician. Diagnostic performance metrics were calculated.
Main Results:
- The AI model accurately generated images for lipedema Types I, II, and III (sensitivity = 1.00).
- Specificity for Type III was 0.50, as images for Types IV and V were misclassified as Type III.
- The model failed to generate any images for Type IV (arm-predominant) or Type V (calf-isolated) lipedema (sensitivity = 0.00). Overall accuracy was 0.600.
Conclusions:
- Generative AI can represent lipedema severity gradients in the lower extremities but fails to depict distinct anatomical subtypes.
- The AI model systematically collapses distinct lipedema subtypes into the Type III phenotype.
- Current generative AI systems may limit reliability for medical education and clinical communication due to their inability to represent lipedema as a distributed anatomical entity.
