Machine-to-Machine Transfer Function in Deep Learning-Based Quantitative Ultrasound.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Quantitative Ultrasound

Background:

• Deep learning (DL) in quantitative ultrasound (QUS) faces challenges with data mismatches across different acquisition systems.
• Existing transfer function methods address acquisition-level mismatches for single scanners.
• Scaling these methods to multiple machines is crucial for broader DL model application.

Purpose of the Study:

• To introduce and evaluate a machine-to-machine (M2M) transfer function for mitigating data mismatches between ultrasound systems.
• To demonstrate the potential for reducing DL model development costs and enabling cross-machine model transfer.
• To assess the impact of calibration phantoms and transducer variations on the M2M transfer function's performance.

Main Methods:

• Developed and applied a novel M2M transfer function to address inter-scanner data variability.
• Utilized two distinct ultrasound machines (SonixOne, Verasonics) with different transducer arrays (L9-4, L11-5).
• Conducted acquisitions using stable and free-hand methods with two calibration phantoms, employing Wiener filtering-inspired implementation.

Main Results:

• Without the M2M method, cross-system classification accuracy was 50% (AUC 0.405).
• With the M2M method, mean accuracy increased to 99% (AUC 0.999).
• Calibration phantom choice significantly impacted performance; the method proved effective with different transducers and a single calibration view.

Conclusions:

• The proposed M2M transfer function effectively mitigates data mismatches at the machine level in DL-based QUS.
• This approach significantly enhances cross-system model generalizability and reduces development costs.
• The method offers a robust and efficient solution for transferring DL models between diverse ultrasound hardware.