Evaluation of generative artificial intelligence in producing anatomically distinct lipedema subtypes: A diagnostic accuracy study

Introduction

Lipedema is a chronic and progressive adipose tissue disorder that primarily affects women. It is characterized by bilateral, disproportionate accumulation of subcutaneous fat and is commonly associated with pain, tenderness, easy bruising, and a sensation of heaviness in the affected limbs, all of which can substantially impair quality of life.^1–3 Although it is estimated to affect approximately 11% of the female population, lipedema remains underrecognized and is frequently mistaken for obesity, lymphedema, or chronic venous disease.^4–6 Accurate diagnosis and recognition of its clinical features are therefore essential for appropriate patient management.

Because the clinical presentation of lipedema varies considerably among patients, several classification systems have been proposed to describe disease distribution and severity. According to the Schmeller classification, Type I is confined to the hips and pelvic region, Type II extends from the hips to the knees, Type III involves the entire lower extremity, Type IV affects the upper extremities, and Type V is limited to the lower legs.^7,8 In addition to anatomical distribution, lipedema is categorized into four clinical stages based on structural changes in the skin and subcutaneous tissue. Early stages are characterized by relatively preserved skin contours, whereas advanced stages may show prominent fat lobules, tissue deformities, and associated lymphatic involvement.^9,10

Advances in generative artificial intelligence have increasingly influenced the creation of medical images. Text-to-image models such as DALL·E and Stable Diffusion, as well as image-generation tools integrated into conversational AI systems, are now widely used to produce educational illustrations, presentation materials, patient information resources, and social media content.^11–13 Despite their growing popularity, concerns remain regarding the accuracy of these images. Since model outputs are shaped by patterns present in the training data, certain patient groups or clinical characteristics may be underrepresented or portrayed inaccurately. Previous studies have highlighted potential biases related to demographic representation and anatomical fidelity.^14–16

These concerns may be especially relevant for conditions with limited visual documentation. When authentic clinical photographs are scarce, AI-generated images can play a disproportionate role in shaping how patients, trainees, and healthcare professionals perceive a disease. ¹⁷ Lipedema represents a notable example. Clinical images depicting Type IV (upper-extremity predominant) and Type V (isolated lower-leg involvement) disease are relatively uncommon in both the published literature and online image repositories, despite the fact that these phenotypes constitute a meaningful part of the clinical spectrum.^8,18 Whether contemporary generative AI systems can accurately depict these less frequently represented subtypes, or instead default to the more familiar lower-extremity phenotype, remains largely unexplored.

Recent diagnostic-accuracy studies of large language and multimodal models in medical image tasks have produced mixed results. For example, a recent evaluation of ChatGPT-4o for grading knee osteoarthritis radiographs found low sensitivity across Kellgren–Lawrence grades and an overall accuracy of 0.230. ¹⁹ Similar diagnostic-accuracy frameworks have been applied to dermatology, ²⁰ ophthalmology, ²¹ and radiology more broadly. ²² To our knowledge, no published study has applied this framework in the reverse direction—generation rather than classification—to evaluate whether AI-produced images of a clinical condition faithfully represent its anatomical subtypes.

The aim of this study was to evaluate the diagnostic accuracy of a widely used generative AI model in producing images consistent with each of the five lipedema types defined by the Schmeller classification, using a 5 × 5 confusion matrix and standard diagnostic metrics. Our primary hypothesis was that the model would accurately depict the dominant lower-extremity types (I–III) but would fail to produce anatomically distinct images for arm-predominant (Type IV) and calf-isolated (Type V) disease.

Methods

Results

A total of 300 images were generated and all were evaluable (0% exclusion). The model produced an interpretable image for every prompt; no safety refusals occurred. The independent reviewers reached consensus on 297 of 300 images (99.0% raw agreement); the three discordant images were adjudicated by the third reviewer.

The confusion matrix of requested versus observed type is shown in Figure 3. The model produced anatomically consistent images for the three lower-extremity types (Type I, II, and III). Sixty of 60 images requested as Type I were classified as Type I, 60 of 60 images requested as Type II were classified as Type II, and 60 of 60 images requested as Type III were classified as Type III. In contrast, the model failed entirely to produce images consistent with Type IV (arm-predominant) or Type V (calf-isolated) lipedema. All 60 images requested under the Type IV prompt and all 60 images requested under the Type V prompt were classified as Type III by the independent reviewers, representing a complete misclassification of these subtypes onto the dominant lower-extremity phenotype.OPEN IN VIEWER

Per-type diagnostic performance is summarized in Table 1 and Figure 4. Sensitivity, specificity, PPV, NPV, and F1-score were all 1.00 for Types I and II. Type III showed perfect sensitivity (1.00) but reduced specificity (0.50) and PPV (0.33) due to the systematic over-production of Type III images when other types were requested. Types IV and V showed sensitivity and PPV of 0.00, with retained specificity of 1.00 (no false-positive generation of these phenotypes in other prompts).

OPEN IN VIEWER

Table 1.

Per-type diagnostic performance metrics of the generative AI model in producing anatomically distinct lipedema subtypes (n = 60 images per type; total n = 300).

Type	Sensitivity	Specificity	PPV	NPV	F1-score	AUC
Type I	1.00	1.00	1.00	1.00	1.00	1.00
Type II	1.00	1.00	1.00	1.00	1.00	1.00
Type III	1.00	0.50	0.33	1.00	0.50	0.75
Type IV	0.00	1.00	—	0.80	0.00	0.50
Type V	0.00	1.00	—	0.80	0.00	0.50
Overall	Acc 0.60	κ 0.50	wκ 0.67	​	Micro 0.75	Macro 0.75

OPEN IN VIEWER

Figure 4.

Per-type diagnostic performance metrics for the generative AI model. Bars represent sensitivity, specificity, PPV, and F1-score for each lipedema type (n = 60 images per type). Sensitivity and PPV reach 1.00 for Types I, II, and III but fall to 0.00 for Types IV and V. Specificity remains 1.00 for Types I, II, IV, and V but drops to 0.50 for Type III. PPV and F1-score are undefined for Types IV and V (no true positives) and shown as 0. Dashed line indicates 0.50 reference.

Overall accuracy across all five types was 0.600 (180/300, 95% Wilson CI 0.543–0.654). Unweighted Cohen κ between requested and observed type was 0.500, and quadratic-weighted Cohen κ was 0.667. Receiver operating characteristic curves for each type, together with micro- and macro-averaged curves, are shown in Figure 5. Type-specific AUC values were 1.00 (Type I), 1.00 (Type II), 0.75 (Type III), 0.50 (Type IV), and 0.50 (Type V). Micro- and macro-averaged AUCs were both 0.75.OPEN IN VIEWER

The per-type metric profile is visualized in Figure 5. The dichotomy between the three lower-extremity types (Types I–III) and the two anatomi1cally distinct subtypes (Types IV and V) is striking: while Types I, II, and III show high sensitivity, PPV, and F1-score, Types IV and V show complete failure across all metrics except specificity, which is preserved because no images of other prompts were misclassified as Type IV or Type V.

Discussion

In this prospective audit of 300 generative AI images, we found a striking dissociation between the model's ability to reproduce intra-type severity within lower-extremity lipedema and its ability to depict anatomically distinct subtypes. The model produced anatomically faithful images for Types I through III but failed entirely to generate images consistent with Type IV (arm-predominant) or Type V (calf-isolated) lipedema; all 120 images requested under these prompts were classified as Type III. This pattern—high fidelity for the dominant lower-extremity phenotype and complete collapse for anatomically distinct subtypes—suggests that the model has learned lipedema as a single visual phenotype rather than as a distributed anatomical entity.

Several mechanisms may underlie this observation. First, the distribution of training data is likely heavily skewed toward Type III lipedema, which dominates online clinical atlases, patient advocacy materials, and indexed photographic repositories.^13,18 Type IV (arm-predominant) lipedema, in particular, is poorly represented in publicly available image collections, despite constituting an estimated 10–30% of clinical cases depending on cohort definitions. ⁸ Second, the semantic embedding of “lipedema” within the model may be anchored to a leg-centric prototype, such that numerical type qualifiers (“Type IV”) act as severity dials rather than anatomical redirectors. Third, even when the model is explicitly instructed to depict arm involvement, the visual prior of lower-extremity disease may dominate the latent representation, a phenomenon previously described in text-to-image bias studies.^14,15

Our findings have direct clinical implications. Lipedema is already widely underdiagnosed, with a median time from symptom onset to diagnosis exceeding a decade in some cohorts.^4,5 Patients with arm-predominant disease in particular are vulnerable to delayed recognition, as the clinical pattern diverges from the popular mental image of lipedema as a “large legs” condition.^8,18 If generative AI is increasingly used to produce illustrative materials for medical education, social media patient outreach, and clinical communication—and current evidence suggests that it is^11–13,17—then the systematic absence of arm-predominant and calf-isolated phenotypes in AI-generated content risks reinforcing the existing diagnostic blind spots. Trainees exposed primarily to AI-generated lipedema imagery may internalize a phenotypically narrow concept of the disease.

These results extend a small but growing literature on bias in generative AI applied to medical content. Prior work has documented underrepresentation of darker skin tones in AI-generated dermatology images, ¹⁶ stereotypical demographic representations in occupation-related prompts, ¹⁴ and inaccuracies in anatomical structures depicted by text-to-image models. ²² Our study contributes an explicit, quantitative diagnostic-accuracy framework—paralleling that used by Temel et al. for ChatGPT-based Kellgren–Lawrence grading ¹⁹ —to evaluate generative rather than discriminative performance. We believe this generative diagnostic-accuracy framework can be applied to other clinical entities with multiple anatomical subtypes (e.g., psoriasis variants, melanoma subtypes, lymphedema stages) to reveal which subtypes are reliably depicted and which are systematically silenced.

The discrepancy between intra-type severity fidelity (Types I–III) and inter-type anatomical fidelity (Types I–III vs IV–V) is itself informative. The model appears to encode a continuous severity axis—producing visibly different images for mild, moderate, and severe lower-extremity disease—but lacks a comparable anatomical axis for redistributing involvement to the upper limbs or restricting it to the lower legs. This dissociation may reflect a more general property of current generative models: continuous visual gradients (severity, age, body size) are learned more readily than discrete anatomical reassignments (which body part is affected), particularly when the discrete categories are rare in training data. ¹⁵

Several limitations should be considered. First, we evaluated a single commercially available image generation interface; results may not generalize to other models or to future versions. Diagnostic-accuracy comparisons across multiple platforms (DALL·E 3, Stable Diffusion, Imagen, FLUX) would be informative and are planned in future work. Second, the prompts used were minimal and did not include synonyms, alternative phrasings, or expanded clinical descriptors; the model's behavior under richer prompting (e.g., “lipedema affecting only the arms”) may differ and warrants separate investigation. Third, two clinicians performed image classification; while consensus was very high (99.0% raw agreement), a larger reviewer panel might detect subtler misclassifications. Fourth, the study evaluated only the anatomical type, not stage, body habitus, demographic characteristics (age, ethnicity), or skin findings; these dimensions of bias are addressed in companion analyses. Finally, the use of synthetic images precludes direct comparison with patient outcomes; clinical impact will depend on downstream use of AI-generated materials in education and communication.

Despite these limitations, the findings are clear: a widely accessible generative AI model accurately depicts the dominant lower-extremity phenotype of lipedema but fails to produce anatomically distinct images of arm-predominant or calf-isolated disease. Clinicians, educators, and patient organizations relying on AI-generated images should be aware of this systematic phenotypic narrowing and should supplement AI output with curated clinical photographs covering the full anatomical spectrum.

Conclusion

A widely used generative AI model produced anatomically faithful images of Types I, II, and III lipedema but failed entirely to generate images consistent with Type IV or Type V lipedema, collapsing all 120 such prompts onto the dominant Type III lower-extremity phenotype. Overall diagnostic accuracy was 0.60, with unweighted Cohen κ of 0.50 and quadratic-weighted Cohen κ of 0.67. The model appears to encode lipedema as a single visual phenotype rather than as an anatomically distributed clinical entity. These findings have implications for the use of generative AI in medical education, patient communication, and clinical illustration of underrecognized anatomical subtypes of lipedema, and motivate a broader generative diagnostic-accuracy framework for other conditions with multiple anatomical variants.