An Ultrasound Imaging Repository of the Pediatric Pancreas: Quantitative Image Feature Analysis and Preliminary Feasibility Assessment of Image-Guided Therapeutic Ultrasound Diabetes Treatment.

Introduction There is a lack of comprehensive pediatric pancreatic imaging repositories, which limits ultrasound (US)-based diagnosis and development of potential interventions, such as TUS stimulation for insulin release in diabetes mellitus (DM) therapy. In light of this, a US imaging repository of the pancreas was created to assess diagnostic and therapeutic ultrasound (TUS) feasibility in pediatric patients with DM. US imaging features were extracted to distinguish between DM and non-DM pancreatic tissue, informing preliminary evaluation of patient-specific TUS models. Methods Clinical records of pediatric patients were reviewed for the presence of pancreatic US imaging and DM diagnosis. Abdominal US images were reviewed qualitatively with a board-certified radiologist for the presence of the pancreas. Upon pancreas identification, image processing was used to extract quantitative features from the pancreas to compare between DM and non-DM patients. Statistical evaluation was performed via Student's t-test, Spearman's correlation (ρ), mutual information (MI) score, and univariate logistic regression to assess feature predictive performance. Receiver operating characteristic (ROC) curves and area under the curve (AUC) values were calculated for each feature, with 95% confidence intervals (CIs) estimated via stratified bootstrapping (n = 2,000 resamples) to quantify variability. Preliminary TUS modeling from the analyzed images was performed to confirm thermal safety effects of potential TUS protocols for DM therapy. One Food and Drug Administration (FDA)-approved high-intensity therapeutic ultrasound software and one previously published TUS model were validated against an in-house Python TUS simulator informed by patient-specific images. Root mean square error (RMSE) was calculated between the thermal outputs generated by each simulation software. Results A US repository for 11 DM and 11 non-DM pediatric patients was created. Imaging features such as entropy (Spearman coefficient (ρ): -0.57, p<0.001, MI score: 0.27), and Centroid Y (ρ: -0.61, p<0.001, MI score: 0.31) support distinctive features in DM patients. Univariate logistic regression analysis identified LBP energy (AUC = 0.79, 95% CI: 0.70-0.87), Haralick inverse difference moment (AUC = 0.78, 95% CI: 0.69-0.86), and Haralick angular second moment (AUC = 0.78, 95% CI: 0.69-0.86) as high performing predictors of DM status. Bootstrapped CIs highlighted stable predictive performance for these texture uniformity features, whereas other descriptors, such as Centroid X and gradient magnitude, showed lower discrimination with CIs falling below the threshold for random predictive performance (0.5). For TUS-mediated insulin release, focused and unfocused simulations confirm safe targeting of pancreatic tissue accessible to treatment with TUS at 1 MHz, 5W/cm