Does quanitative ultrasonographic analysis performed with artificial intelligence methods contribute to the diagnosis of lipedema? a pilot preliminary study

Objective: Lipedema, a chronic disorder of subcutaneous adipose tissue, is common yet often overlooked in clinical practice and frequently mistaken for obesity. Ultrasonographic (US) imaging methods beyond clinical examination have not been sufficiently studied in the differential diagnosis. The aim of this study is to investigate whether quantitative, AI‑assisted radiomics analysis of lower‑limb US images adds diagnostic value for lipedema. Methods: The study included ultrasonographic scans from four women with clinically confirmed lipedema and four age-matched women with obesity (BMI ≥ 30 kg/m²) referred to Dokuz Eylül University, Faculty of Medicine, Department of Physical Therapy and Rehabilitation outpatient clinic with suspected lipedema. All examinations were performed in the outpatient musculoskeletal US unit by an experienced radiologist. Using fixed depth, sector width and gain settings, eight images per leg were obtained at four standardised sites (2): mid‑anterior thigh, pretibial mid‑shaft, mid‑lateral leg, and medial supramalleolar region (Fig. 1). Also dermal and subcutaneous thicknesses were recorded. The optimal image area (region of interest = ROI) was labeled in these retrospective images, and radiomics analysis (tissue texture) was performed in the system. For this purpose, the LIFEx program (www.lifexsoft.org) was used. With this method, various pattern features were extracted from the determined ROIs. Dimensionality reduction techniques such as feature selection or Principal Component Analysis (PCA) were used to reduce the data size, which increased with the extracted features. The resulting dataset was analyzed for clinically significant results in images with lipedema. Given the small sample (n = 4 per group), group differences were evaluated with an independent‑samples t‑test using sample bootstrapping (100 resamples). Results: Radiomics features were compared between groups using a bootstrapped t-test. Within the intensity‑based category, Mean Intensity (p < 0.001) and Intensity Skewness (p = 0.01) differed significantly between lipedema and obesity images. From the grey‑level co‑occurrence matrix (GLCM), both Contrast (p < 0.001) and Dissimilarity (p = 0.02) showed significant difference. In addition, the grey‑level run‑length matrix (GLRLM) feature High Grey Level (p = 0.003) and the grey‑level size‑zone matrix (GLSZM) feature Grey Level Variance (p = 0.03) were statistically distinct between groups. No statistically significant differences were detected in dermal or subcutaneous thicknesses, except for dermis thickness at site A on the right limb, which reached significance (p = 0.04). Conclusion: This preliminary study suggests that quantitative ultrasonographic analysis enhanced by artificial intelligence–based radiomics may provide supportive diagnostic value in distinguishing lipedema from obesity. Larger-scale studies are needed to validate these preliminary findings. Keywords: Lipedema, Obesity, Radiomics Analysis, Artificial Intelligence