Using Ultrasound Image Augmentation and Ensemble Predictions to Prevent Machine-Learning Model Overfitting

Area of Science:

• Medical imaging
• Artificial intelligence
• Diagnostic tools

Background:

• Deep learning models offer potential for automating medical imaging diagnostics.
• Generalizability across subject variability is crucial for clinical application.
• Previous ShrapML model showed <70% accuracy for shrapnel identification.

Purpose of the Study:

• To improve the test accuracy and generalizability of a deep learning image classifier for shrapnel identification.
• To evaluate the impact of data augmentation and ensemble methods on model performance.

Main Methods:

• Utilized a neural network (ShrapML) with leave-one-subject-out (LOSO) cross-validation.
• Implemented image augmentation techniques (affine transformations) and MixUp for generating training data.
• Employed ensemble prediction strategies, including bagging confidences/predictions from multiple LOSO holdouts.

Main Results:

• Initial ShrapML accuracy was <70%, with variability.
• Data augmentation and MixUp improved accuracy to 75%.
• Top-3 LOSO confidence bagging achieved >85% accuracy on blind tissue phantoms.
• Gradient-weighted class activation mapping confirmed shrapnel tracking.

Conclusions:

• Data augmentation and ensemble prediction approaches enhance the generalization of predictive models for ultrasound image analysis.
• These methods are critical for developing reliable, real-time diagnostic deployment systems.
• Improved accuracy and reliability are key for clinical adoption of AI in medical imaging.