Detecting Uniformity Artifacts in Ultrasound Transducers

Abstract

Objectives

Uniformity artifacts caused by defective transducer elements or scanner malfunctions degrade diagnostic image quality. Traditional quality control (QC) methods, such as phantom testing and visual or quantitative image analysis, can be labor‐intensive and limited in test frequency. This study aims to develop a deep learning framework to detect uniformity artifacts and complement traditional QC.

Methods

Clinical median images were generated by aggregating co‐registered grayscale ultrasound images acquired by each transducer and computing median intensity values across the image stack. A pretrained ResNet‐18 model was fine‐tuned on a dataset consisting of clinical median images from linear and curvilinear transducers. The dataset was divided into training, validation, and testing subsets, ensuring no overlap between training and test transducers. To assess generalizability, the model was also evaluated on an independent test set of 396 phantom‐validated images from linear and curvilinear transducers.

Results

The model achieved 100% accuracy on the first dataset's test set. On the independent test set, it attained 87.4% accuracy with high sensitivity (0.84) and specificity (0.88), demonstrating robust generalization. Occlusion sensitivity maps confirmed the model's attention to uniformity artifact regions.

Conclusion

The deep learning framework using clinical median images demonstrated robust performance across several linear and curvilinear transducer models. It could be integrated into the clinical QC workflow by automating artifact detection in an effective and timely manner. In our practice, it can flag median images classified as artifact‐present, reducing human review time by approximately 80% while preserving detection accuracy.

Abbreviations

CNN

convolutional neural network

QC

quality control

SDR

systematic dark region

Uniformity artifacts are the most common findings in quality control (QC) testing of ultrasound systems. ¹ These artifacts may appear for a variety of reasons including weak or dead elements, scanner malfunctions, faulty electrical connections, or physical damage to components. These artifacts can adversely impact the quality of grayscale and color Doppler images and undermine the accuracy of velocity and other quantitative measurements. ² , ³ The most common traditional QC methods for detecting these artifacts typically involve periodic testing using with in‐air reverberation, the paper‐clip test, or preferably phantom‐based evaluation. They require physical access to the scanner and transducers, as well as visual or quantitative assessment of the images. ⁴ These traditional QC tests require significant resources from both equipment and personnel so they are typically limited to semi‐annual or annual assessments, allowing defects to potentially go undetected for extended periods of time.

Recently, there has been growing interest in using clinical patient images to detect uniformity artifacts. This approach leverages the observation that defects in the transducer or scanner tend to produce consistent patterns in the images, which can be identified through the analysis of a large number of clinical images. Our team has previously demonstrated the feasibility of detecting uniformity artifacts using median images reconstructed from clinical grayscale images from each transducer. ⁵ This approach is being studied in our clinical practice as a way to enhance transducer artifact detection. Moreover, studies have shown the feasibility of further applying novel methods to process clinical uniformity images for the automatic detection of uniformity artifacts. For instance, Lorentsson et al ⁶ conducted a retrospective study that analyzed over 31,000 clinical images collected from linear transducers from 3 ultrasound scanners and applied their algorithm to detect uniformity artifacts. The results were then compared with the electrical measurements of the transducers when they were known to be defective. Their algorithm showed good visual agreement with the electrical measurements, suggesting that clinical images could be used for automatic detection of defective transducers. The same group ⁷ also proposed a method using systematic dark region (SDR) curves to detect defects in clinical images, expanding the scope to include both linear and curvilinear transducers and comparing the method's performance against the in‐air reverberation method.

This work is based on our clinically implemented semi‐automated workflow, which aggregates registered grayscale ultrasound images acquired by each transducer, geometrically aligns images to account for different imaging depths, computes median intensity values across the image stack to suppress anatomical noise, and flags artifacts via expert review. While this semi‐automated method has demonstrated high accuracy in detecting uniformity artifacts and improved artifact identification in our institution, its reliance on manual interpretation poses scalability challenges in clinical environments with numerous transducers. Deep learning has emerged as a transformative tool in biomedical engineering, with demonstrated success in diverse applications spanning genetic research, ⁸ optimizing radiation therapy planning, ⁹ , ¹⁰ , ¹¹ and medical image analysis. ¹² , ¹³ , ¹⁴ These developments highlight its versatility and transformative potential in healthcare and biomedical applications, reaffirming the growing prominence of computational methods in these fields. Building on this momentum, we propose augmenting the pipeline with a deep learning model for automatic uniformity artifact detection in clinical images. The goal is to reduce reviewer burden by analyzing clinical median images from all transducers and excluding those with a very low probability of having an actionable artifact from manual or visual interpretation. By integrating human expertise with automated analysis, this hybrid approach aims to standardize artifact detection while accelerating workflow throughput, which is a critical step toward scalable transducer QC in high‐volume clinical settings. The goal of this work is to enable early and automated detection of uniformity artifacts and complement traditional QC.

Materials and Methods

Datasets and Image Processing

In our clinical practice, custom MATLAB (MathWorks, MA) programs were used to create clinical uniformity images on a periodic basis by identifying the most recent 200 grayscale images from each unique transducer and scanner combination. Figure 1 shows the flowchart of our current image processing steps used to generate clinical median images for each transducer. A custom MATLAB program would collect clinical grayscale DICOM images from the past 8 weeks for each transducer (Figure 1, A–C). The text and informational part of the image were cropped out and only the image area was kept (Figure 1D). If color images are found (annotations, etc.), RGB would be converted to grayscale (Figure 1E). For GE LOGIQ E9 scanners (GE Healthcare, WI), clinical grayscale DICOM images were sorted according to transducer model and scanner inventory (Figure 1F). Images were then anonymized, segmented, scaled to a consistent pixel size, and co‐registered (Figure 1G). Segmentation was applied over a consistent physical region defined by the full field of view and a predefined depth for each probe. Co‐registration was performed using a rigid transformation. Any pixels outside the active image area for a given image, due to smaller FOV or insufficient depth, were assigned NaN values. We also removed annotations and set pixel groups with over 500 connected components having grayscale values at or near 0 to NaN (Figure 1, H and I). Four median images were generated from sets of 25, 50, 100, and 200 images by calculating the median grayscale value for each pixel in the co‐registered images to capture the temporal trend (Figure 1, J and K). If a transducer has fewer images than the targeted quantity of images, a median is instead computed from the available images.

Figure 1.

Flowchart of the processing steps used to generate clinical median images for each transducer.

Two distinct datasets were employed to develop and evaluate the artifact detection model. This retrospective study used de‐identified clinical images. It was reviewed by the local Institutional Review Board and granted exempt status. The primary dataset ¹⁵ comprises images acquired using both linear transducers (9 L‐D, L8‐18i, ML6‐15‐D) and curvilinear transducers (C1‐5‐D, C1‐6‐D, C2‐9‐D) from GE LOGIQ E9 scanners, collected over a 4‐month period. These images were labeled by a single experienced reader based on semi‐annual and annual QC records. For linear transducers, 241 artifact‐free median images from 45 transducers and 228 artifact‐containing images from 14 transducers were selected. Similarly, for curvilinear transducers, 263 artifact‐free median images from 35 transducers and 256 artifact‐containing images from 17 transducers were identified. To enhance the model's robustness, the dataset underwent augmentation via translation and mirroring techniques. It was then partitioned into training, validation, and testing sets using a transducer‐level split. Specifically, each transducer produced multiple images, and the test set included only images from a distinct subset of transducers that were entirely excluded from the training and validation sets. This ensured independence between sets and prevented data leakage. The training and validation images were randomly sampled from the remaining transducers. The proportion of images from each transducer model was kept about the same in each image set. The final distribution included: 22,226 training images (86.5%), 2,162 validation images (8.4%), and 1300 testing images (5.1%).

A second independent dataset, ⁴ separately labeled by 2 experienced readers and notably utilizing phantom testing results retrospectively from QC reports as the ground truth, was reserved for external validation. This dataset contains 396 median images collected from 108 transducers including both linear transducers (9L‐D, L8‐18i, ML6‐15‐D) and curvilinear transducers (C1‐5‐D, C1‐6‐D, C2‐9‐D) from GE LOGIQ E9 scanners and was used exclusively to evaluate the model's robustness and stability. By incorporating phantom‐validated labels and multi‐reader consensus, this external test set ensures rigorous evaluation of generalizability. For both datasets, only image‐level labels were provided, and no annotations indicated the exact locations of artifacts within the images.

Model Architecture and Training Protocol

To leverage pre‐existing feature representations and improve training efficiency, ¹⁶ we adopted a transfer learning approach using ResNet‐18, ¹⁷ a pretrained convolutional neural network (CNN), which was selected due to its proven effectiveness in medical image analysis and its ability to capture hierarchical features while maintaining a relatively low computational cost (Figure 2). ¹⁸ We broadcasted our image dimension (batch, 1, 224, 224) to (batch, 3, 224, 224) to match with the original input configuration, replicating each 1‐byte/pixel grayscale intensity value across the 3 color channels to form a 3‐byte/pixel grayscale image. To tailor the model for binary artifact detection (artifact versus no artifact), the final fully connected layer of ResNet‐18 was replaced with a custom fully connected layer (2 neurons) followed by a softmax activation layer. The Adam optimizer was employed with an initial learning rate of 0.0005 to balance stability and convergence speed. A piecewise learning rate schedule reduced the rate by a factor of 0.2 every 5 epochs, allowing gradual refinement of weights during training. ¹⁹ , ²⁰ Training proceeded for 15 epochs with a mini‐batch size of 64, and data was shuffled every epoch to minimize batch‐order bias. Validation metrics were evaluated every 30 mini‐batches (validation frequency = 30) to monitor generalization performance. Occlusion sensitivity maps with a mask size of 10 × 20 ¹⁵ were plotted on the positive results from the test set to identify the pixel areas that contributed to the classification results.

Figure 2.

ResNet‐18 architecture. ¹⁷

Results

The model exhibited exceptional image‐wise performance on the primary dataset, achieving 100% accuracy on the training set and 95.8% accuracy on the validation set by the final epoch. Crucially, it retained 100% accuracy on the test set, which consisted exclusively of images from unique transducers withheld during training and validation, demonstrating robust generalization within the primary dataset.

When evaluated on the independent dataset ⁴ with phantom testing results retrospectively from QC reports as ground truth, the model achieved 87.4% image‐wise accuracy (95% CI: 83.7–90.3), with a sensitivity of 0.84 (95% CI: 0.73–0.91) and specificity of 0.88 (95% CI: 0.84–0.91), respectively.

Figure 3 shows the occlusion sensitivity maps overlaid on the clinical median images for 2 example probes with artifacts, identifying pixel areas which contributed to the positive results. These occlusion sensitivity maps confirm the model's focus on artifact regions, aligning with human interpretation.

Figure 3.

Examples of occlusion maps overlaid on the clinical median images with artifacts. Top: clinical median image calculated from 50 clinical images on a 9 L‐D transducer; Bottom: clinical median image calculated from 200 clinical images on a C1‐6‐D transducer. The regions on both sides appear different because the clinical images may have had varying fields of view (FOV) during alignment. To register them into the same space, some pixel regions are null in images with smaller FOVs.

Discussion

In this study, our approach demonstrated exceptional performance, achieving highly accurate results based solely on a deep learning model developed on single median images. We specifically selected ResNet‐18 after comparing its performance with deeper CNNs, including ResNet‐101. Both models were trained and validated under the same experimental conditions and evaluated on the same test set. ResNet‐101 achieved 100% accuracy for linear transducers and 98.6% for curvilinear transducers, ¹⁵ whereas ResNet‐18 achieved 100% accuracy for both. These results demonstrate that increasing model depth and complexity did not yield additional performance gains for this dataset and may increase the risk of overfitting. In addition, ResNet‐18 contains approximately 11 million trainable parameters compared with about 44 million in ResNet‐101, which means it requires significantly less computational resources.

Our prior reader study which also utilized images from the independent dataset, ⁴ evaluated 111 transducers from 44 GE LOGIQ E9 scanners and employed the same clinical median image processing pipeline as the current work. A custom MATLAB program aggregates registered ultrasound images, computes median intensity values, and generates a standardized 4‐quadrant “summary image” (composed of 25‐, 50‐, 100‐, and 200‐image medians) (Figure 4). Artifacts were assessed by 2 readers with >10 years of experience in ultrasound performance testing, ensuring methodological consistency across studies. Across 8 transducer models and 115 test events, the clinical uniformity method demonstrated high specificity (0.99) and sensitivity (0.78) for detecting any artifact, whether subtle or significant. Furthermore, in our clinical practice, we prioritize fixing artifacts that we judge to have significant impact on image quality and diagnostic accuracy, and refer to these as “actionable artifacts.” This is a subjective assessment that considers factors including conspicuity of the artifact in single frame grayscale images of real anatomy, the location and number of artifacts present, and the specific needs of the clinical practice area. Different practices may well have different criteria for actionable artifacts. When focusing solely on clinically actionable artifacts, the sensitivity increased to 0.88, but the specificity decreased to 0.93 due to false positives caused by subtle artifacts that did not require transducer replacement. The 2 experienced observers agreed on all assessments of clinical uniformity images, with an average read time of approximately 11 seconds per image. The clinical uniformity method successfully detected artifacts in 14 true‐positive cases for any artifact and 7 true‐positive cases for clinically actionable artifacts. However, it produced 1 false‐positive result for any artifact and 8 false positives for clinically actionable artifacts, which could lead to unnecessary physical verification efforts. The method missed 4 false‐negative cases for any artifact and only 1 for clinically actionable artifacts, the latter being an intermittent defect. The results demonstrated great potential for using clinical median images to detect ultrasound uniformity artifacts and equipment defects, but a certain amount of review time per image was still required.

Figure 4.

An example of a “summary” image from reference. ⁴ The reviewer evaluated each transducer based on its “summary” image, while the deep learning model used individual quadrant images for training, validation, and testing.

As a comparison, we also adapted our deep learning framework, which generates predictions per quadrant image, for transducer‐level assessments by applying the decision protocol established in our prior clinical workflow. ⁴ Specifically, artifacts need to show consistency across 25‐, 50‐, and 100‐image quadrants to flag the transducer. To validate performance, we compared the model's output with the results from 2 experienced readers on the independent test set, as well as with the phantom uniformity test results, which served as the ground truth. ⁴ The phantom images were visually assessed by QC. Analysis of model outputs against phantom assessments revealed that, within the independent test set, most errors occurred in edge cases that typically fall outside the scope of the model's training data or involve conditions of greater uncertainty. All the false‐negative predictions were due to artifacts that are not clinically actionable, and both readers also labeled them as artifacts. Among 39 false‐positive image‐wise predictions, 22 stemmed from preprocessing failures (eg, low quality clinical median images, residual text overlays). Of the remaining 17, 7 showed artifacts confined to a single quadrant of the transducer's “summary” image.

If we applied our prior clinical criteria where artifacts require consistency across 25‐, 50‐, and 100‐image quadrants, only 4 transducers would be misflagged as positive. Across all 108 test events, the probe‐wise accuracy was 96.3%, with high specificity (0.96) and sensitivity (0.82). Compared to reference, ⁴ which used the same dataset, we excluded the S1‐5 and S4‐10 models in this study, as they were not part of the primary dataset. Additionally, 1 probe was excluded due to the intermittent artifact caused by cable flex, which could not be detected in the uniformity images. Our model achieved slightly higher sensitivity, while the accuracy and specificity remained nearly the same. Notably, all the 3 false‐negative predictions were due to subtle and non‐actionable artifacts. Neither our model nor the 2 readers were able to identify these artifacts. Among the 4 false‐positive transducers, 1 was also mislabeled by both readers, but was later confirmed as negative through phantom scans, making it the sole false positive in the readers' assessments. When focusing solely on significant artifacts, the probe‐wise accuracy will be 92.6%, with increased sensitivity (1) and decreased specificity (0.89). Overall, our model was able to produce prediction results that were almost the same as those of the 2 experienced readers for these subtle artifact cases. If our model had been used to prescreen the images in the independent test set, it could have reduced the number of images requiring human review to just 16.7%, without missing any clinically actionable artifacts.

These results highlight the model's clinical relevance and stability when faced with interference or unseen data, while emphasizing its per‐image diagnostic consistency across diverse acquisition conditions. By aligning the model's output with prior clinical decision‐making protocols, where multi‐quadrant consistency is required, the practical impact of false positives is minimized. Moreover, by flagging only suspicious cases, the model significantly cuts review time compared to reviewing all images. The results showed that the model replicated reader performance, including similar patterns in subtle artifact cases, while automating detection to reduce manual review time without compromising accuracy.

In addition, this algorithm was also recently tested using a dataset based on images from the GE LOGIQ E10 scanner platform ²¹ and achieved results with an accuracy of 87.3%, sensitivity of 0.93, and specificity of 0.87. With the level of training and performance reported here, we believe this model could be used in clinical practice to enhance our ability to detect actionable artifacts more effectively than our current semi‐annual clinical QC testing process. The process could be currently applied on both linear transducer models (9L‐D/L2‐9‐D, L8‐18i, ML6‐15‐D) and curvilinear transducer models (C1‐5‐D, C1‐6‐D, C2‐9‐D) from GE LOGIQ E9 and E10 ²¹ scanners. The proposed deep learning‐based method can serve as a pre‐screening tool within the existing QC workflow in our practice. Given its high sensitivity for detecting potential uniformity artifacts, the model can efficiently identify images that may need further review while excluding those that are artifact‐free. This approach enables physicists to focus on a smaller subset of flagged cases, thereby improving workflow efficiency and reducing review time. For the flagged images, physicists can first review them to exclude false positives and assess the severity of the detected artifacts, followed by targeted phantom uniformity tests on relevant transducers. The estimated human time cost could be reduced by approximately 80%. Future work will focus on automating the detection of actionable artifacts, and extend this approach to other practices, thereby generating more data to support a standardized method.

This study has several limitations. First, the median image approach is not as sensitive as robust phantom QC testing due to the clinical use of spatial compounding and image processing, and it may not be able to capture intermittent artifacts. Second, despite implementing preprocessing methods, residual text overlays persisted in some median images, contributing to classification errors. Developing robust text detection algorithms or expanding preprocessing steps could address this issue. Third, the model was trained exclusively on GE LOGIQ E9 scanners and 6 transducer models. Expanding the training dataset to incorporate other scanner platforms and additional transducer types would enhance generalizability and clinical applicability.

Conclusions

Our ResNet‐18‐based model enables efficient, automated detection of uniformity artifacts across different datasets with varying types of transducers, achieving performance comparable to human reviewers. By reducing manual review time and maintaining high accuracy, this approach enhances QC workflows and supports timely defect resolution. This advancement lays the groundwork for more effective artifact detection in routine practice.

Data Availability Statement

The datasets generated or analyzed during this study are not publicly available. Requests for access to the data should be directed to the corresponding author.