Calibrating Data Mismatches in Deep Learning-Based Quantitative Ultrasound Using Setting Transfer Functions.

Area of Science:

• Medical imaging
• Artificial intelligence
• Signal processing

Background:

• Deep learning (DL) models in ultrasound imaging are susceptible to performance degradation due to data mismatches between training and testing distributions.
• Acquisition-related data mismatches, stemming from variations in scanner settings, pose a significant challenge for the clinical implementation of DL-powered ultrasound.
• Mitigating these mismatches is essential for the reliable application of DL in ultrasound imaging and tissue characterization.

Purpose of the Study:

• To propose an inexpensive and generalizable method for calibrating data mismatches in ultrasound imaging caused by different scanner settings.
• To enhance the performance of deep learning models in ultrasound classification tasks despite variations in acquisition parameters.
• To enable wider clinical adoption of DL-powered ultrasound by improving model robustness.

Main Methods:

• A large training dataset was collected at a single setting, complemented by small calibration sets at each distinct scanner setting.
• A signals and systems approach was employed to calibrate data mismatches using calculated setting transfer functions.
• Two convolutional neural networks (CNNs), ResNet-50 and DenseNet-201, were trained and evaluated on ultrasound radio frequency (RF) data from phantom classification tasks.

Main Results:

• Without calibration, CNNs achieved mean classification accuracies of 52% (pulse frequency), 84% (focus), and 85% (output power) for different data mismatches.
• After applying setting transfer functions for calibration, mean classification accuracies significantly improved to 96% (pulse frequency), 96% (focus), and 98% (output power).
• The proposed method demonstrated generalizability across three types of data mismatches: pulse frequency, focus, and output power.

Conclusions:

• Incorporating setting transfer functions provides an economical and effective means to generalize DL models for ultrasound classification tasks with variable scanner settings.
• Calibrating for scanner setting mismatches using a signals and systems perspective substantially improves DL model performance and reliability.
• This approach facilitates the wider clinical adoption of DL in ultrasound imaging by ensuring consistent performance across diverse acquisition environments.