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

Abstract

Deep learning (DL) can fail when there are data mismatches between training and testing data distributions. Due to its operator-dependent nature, acquisitionrelated data mismatches, caused by different scanner settings, can occur in ultrasound imaging. As a result, it is crucial to mitigate the effects of these mismatches to enable wider clinical adoption of DL-powered ultrasound imaging and tissue characterization. To address this challenge, we propose an inexpensive and generalizable method that involves collecting a large training set at a single setting and a small calibration set at each scanner setting. Then, the calibration set will be used to calibrate data mismatches by using a signals and systems perspective. We tested the proposed solution to classify two phantoms using an L9-4 array connected to a SonixOne scanner. To investigate generalizability of the proposed solution, we calibrated three types of data mismatches: pulse frequency mismatch, focus mismatch and output power mismatch. Two well-known convolutional neural networks (CNN)s, i.e., ResNet-50 and DenseNet-201, were trained using the ultrasound radiofrequency (RF) data. To calibrate the setting mismatches, we calculated the setting transfer functions. The CNNs trained without calibration resulted in mean classification accuracies of around 52%, 84% and 85% for pulse frequency, focus and output power mismatches, respectively. By using the setting transfer functions, which allowed a matching of the training and testing domains, we obtained mean accuracies of 96%, 96% and 98%, respectively. Therefore, the incorporation of the setting transfer functions between scanner settings can provide an economical means of generalizing a DL model for specific classification tasks where scanner settings are not fixed by the operator.

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

“…Additionally, DL-based QUS approaches which are developed for specific machines, can be transferred to other machines at ease. Overall, in our prior work, we demonstrated that the transfer function approach has potential to provide an economical way to provide in-system transferability [9]. In this work, we demonstrate that transfer functions can be defined that can also provide out-system transferability, i.e., a Machine-to-Machine (M2M) transfer function.…”

“…Specifically, in several recent examples, CNNs were used to classify tissue states, and it was shown that they outperformed traditional QUS approaches [3]- [8]. Following this, in our previous work, a transfer function approach was developed using a calibration phantom to mitigate acquisition-related data mismatches within the same imaging machine for DL-based QUS approaches [9]. The transfer function approach significantly improved mean classification accuracies for pulse frequency, output power, and focal region mismatches within the same imaging machine, increasing them from 52%, 84%, and 85% to 96%, 96%, and 98%, respectively.…”

“…The issue of data mismatch has gained increased attention in recent literature focusing on DL-based QUS [9], [19], [27]- [31]. Adaptive batch normalization was utilized in the context of DL-based QUS [27].…”

“…Furthermore, the Fourier Domain Adaptation technique was employed, proposing the replacement of lower frequency components within the frequency spectrum [30]. In contrast to these methodologies, the transfer function approach developed in [9] does not require real sample data from the testing domain to be used for training. Instead, it relies on a calibration phantom that can be tailored to the specific characteristics of the real sample at hand.…”

“…In a prior work, a transfer function was developed to mitigate acquisition-related data mismatches within the same imaging machine [9]. In this instance, data mismatch occurs at the machine level, including acquisition and hardware-level mismatches.…”