A Co-Plane Machine Learning Model Based on Ultrasound Radiomics for the Evaluation of Diabetic Peripheral Neuropathy.

Area of Science:

• Medical Imaging
• Machine Learning
• Neurology

Background:

• Diabetic peripheral neuropathy (DPN) detection is crucial for preventing severe complications.
• Machine learning (ML) and radiomics show promise for DPN diagnosis, but their use with ultrasound is limited.
• The optimal ultrasound plane (longitudinal vs. transverse) for radiomic feature extraction in DPN remains unclear.

Purpose of the Study:

• To analyze and compare radiomic features from transverse and longitudinal ultrasound planes of the tibial nerve.
• To develop a co-plane fusion ML model for enhanced DPN diagnostic accuracy.

Main Methods:

• 516 feet from 262 diabetics were analyzed across two institutions.
• 1316 radiomic features were extracted from transverse and longitudinal tibial nerve planes.
• Six ML algorithms constructed radiomics models using single and combined planes; performance was evaluated using ROC, calibration curves, and DCA. SHAP analysis explained model predictions.

Main Results:

• The co-plane Support Vector Machine (SVM) model achieved superior AUCs: 0.90 (training), 0.88 (internal testing), and 0.70 (external testing).
• These results significantly outperformed single-plane models (P < 0.05).
• Calibration curves and DCA indicated good model fit and potential clinical utility.

Conclusions:

• The co-plane SVM model, integrating transverse and longitudinal radiomic features, demonstrated optimal performance for DPN prediction.
• This approach significantly enhances DPN diagnostic efficacy.
• The model shows promise as a robust, noninvasive tool for clinical DPN assessment.