Noninvasive identifying of patients with type 2 diabetes by ultrasound wavelet texture features of skeletal muscle

Background: Non-invasive screening for type 2 diabetes (T2D) remains a significant clinical challenge. This study aimed to assess the efficacy of using wavelet texture features from ultrasound images to identify patients with T2D. Methods: This study retrospectively analyzed 292 patients with T2D and 292 matched healthy controls who underwent lower limb ultrasound (June 2023 − July 2025). Controls were selected from routine health screenings, excluding those with medical conditions. Ultrasound images of the gastrocnemius muscle were collected and the wavelet texture features were extracted. Radiomics models for T2D identification were developed and tested using 5-fold cross-validation with multiple machine learning algorithms. Feature selection and model construction are conducted independently within each fold. The models’ final performance was derived from the aggregated results across all folds to ensure generalizability. Results: Our cohort consisted of 392 males and 192 females. Analysis of ultrasound images yielded 833 texture features. The selected features in each fold were utilized to construct classification models. Most models demonstrated fair to good performance for T2D detection, with Logistic Regression (AUC 0.76 ± 0.06, sensitivity/specificity 0.68 ± 0.08 / 0.70 ± 0.06) and Extremely Randomized Trees (AUC 0.76 ± 0.06, sensitivity/specificity 0.67 ± 0.06 / 0.70 ± 0.07) excelling in the test set. Among the features utilized in the development of the LR and ET models, wavelet-HHH GLCM correlation demonstrated a high ranking in both. Conclusions: Machine learning classifiers utilizing wavelet-based textural features extracted from gastrocnemius muscle ultrasound images offer a promising non-invasive screening method for T2D.