Method for automatic detection of defective ultrasound linear array transducers based on uniformity assessment of clinical images — A case study

Abstract The purpose of the present study was to test an idea of and describe a concept of a novel method of detecting defects related to horizontal nonuniformities in ultrasound equipment. The method is based on the analysis of ultrasound images collected directly from the clinical workflow. In total over 31000 images from three ultrasound scanners from two vendors were collected retrospectively from a database. An algorithm was developed and applied to the images, 150 at a time, for detection of systematic dark regions in the superficial part of the images. The result was compared with electrical measurements (FirstCall) of the transducers, performed at times when the transducers were known to be defective. The algorithm made similar detection of horizontal nonuniformities for images acquired at different time points over long periods of time. The results showed good subjective visual agreement with the available electrical measurements of the defective transducers, indicating a potential use of clinical images for early and automatic detection of defective transducers, as a complement to quality control.

Another simple method of detecting reduced sensitivity of transducer elements is to analyze an image with the transducer held in air. 7,20,21 Another known technique for detection of defective transducers is to move a linear transducer along the arm while viewing the dynamic image; it is easy for a human eye to detect vertical streaks in the image when the tissue is moving while the streaks are not. This method, that uses several image frames, is based on the fact that every frame is slightly darker below the defective part of the transducer, since the ability to send and receive echoes is affected regardless if the failure is caused by dead elements, cable failure or delamination in the transducer. In a single image it may be difficult to discriminate these streaks from other details/structures in the image, but by using information from several image frameswhere the image background varies but the streaks remain constant the possibility for detection increases.
Visual inspection of the image uniformity using a tissue-mimicking phantom is used in quality assurance to detect both vertical and horizontal nonuniformities. 3 King et al. 22 compared different ways of detecting defective transducers by using the information in several images of a dynamic clip of a low-cost phantom produced for this purpose. 23 In their study, the median image of the dynamic clip was used in two ways for visual assessment and was compared with assessment of the dynamic clip. All three methods were developed for increased sensitivity for detecting subtle artifacts Electronic transducer testers such as FirstCall aPerio (Sonora Medical Systems, Inc., Longmont, CO, USA) and ProbeHunter (BBS Medical AB, Stockholm, Sweden) are testers to which the transducer is connected. Pulses are sent element-wise toward a target in water and the echoes are evaluated. The result from the test is comprehensive and typically contains, among other parameters, the sensitivity for all elements individually presented as bar graphs. If an element has zero sensitivity, the element is nonfunctional.
The existing methods described above for testing for defective transducers are associated with limitations. Firstly, they all require access to the equipment (or at least the transducer) and are timeconsuming. This requires personal resources for performing the tests and the equipment is furthermore unavailable for clinical use. Secondly, when checking for defects, the time from failure to detection could in worst case be up to the checking interval, and a defect may therefore be unnoticed for a long time. Finally, intermittent failures may not be detected at all by the described methods.
The purpose of the present study was to address the limitations described above by testing an idea of and describing the concept of a novel method of detecting defects related to horizontal nonuniformities in ultrasound equipment. The method is based on the analysis of ultrasound images collected directly from the clinical workflow and applicable to, e.g., linear array transducers, as used in the present study.

2.A | Description of the concept of the method
The proposed method is based on the fact that the anatomical information in an image varies for every clinical image, whereas the darker region corresponding to a defective transducer is present in all images produced by the transducer. Thus, by averaging many images or determining their median, a more or less homogeneous background is obtained even from clinical images, since the anatomical variations tend to cancel each other, whereas the systematic darker streakspresent in all imagesremain and become easier to detect than in a single image.
Basing a method of detecting nonuniformities on clinical images introduces new possibilities for quality control. As the images can be collected directly after they are stored clinically, the quality control can be performed without intervening with the equipment and at any time point. The most straightforward approach seems to be to calculate a median uniformity image for visual assessment based on a number of images that have been stored in a database. In Fig. 1, this is shown for different numbers of images where a transducer with eight defective elements (as determined by a FirstCall measurement) was used (Case 1, see Table 1). An example of a single clinical image acquired with the defective transducer is shown in Fig. 2, indicating that it may be very difficult to detect the defects in the clinical images directly.
For continuous monitoring of the system, a more advanced approach could be to automatically analyze the latest produced images for horizontal uniformity aberrations. Such an analysis could be a complement to the normal quality assurance program and notify the service organization of a detected artifact and thus lead to earlier detection of defects. The details of an example of such an analysis are given in the present paper. The purpose of the paper was to describe the details of this first attempt in a case study based on three ultrasound systems and to test if it may be possible at all. The results are qualitative and a follow-up study is planned for a quantitative evaluation of a larger number of ultrasound systems and cases.

2.B | Application of the method to three general imaging ultrasound scanners
The method for visual assessment and an implementation of an automated analysis was applied to stored images for three scanners,

2.C | Image selection
Images from a certain scanner, distinguished with the DICOM tag "StationName", were retrieved from the archive. The transducer type was determined by the DICOM tag "TransducerData" for the Philips scanner. For the GE scanners, the transducer information in the images was used instead since the DICOM tag was empty for these images. Images containing Doppler curves and side-by-side images were rejected. This was determined by checking if the DICOM tag "SequenceOfUltrasoundRegions" had more than one item or if the size of the image had more columns than the ones just containing one image. Color Doppler images without curves were included.

2.E | Implementation of an algorithm for automatic detection of horizontal nonuniformities
The idea of the automatic analysis was to develop an algorithm that uses the information in the superficial part of a large number of images to create a curve with a length equal to the number of and width of all peaks that were found. The means of the heights and widths of the peaks were calculated. All peaks that had a higher mean value of the height than the threshold T green were selected.
The heights and the widths of the n selected peaks were used to create a vector, containing Gaussian curves for the selected peaks at the correct position, see eq. (1).
For the red and blue paths in Fig. 3, all N stack images were used to calculate a median image. Pixels between rows R upper and R lower were used to calculate the CWM curve, as above. The CWM curve was inverted and for the blue path all values above T blue for the first and last P include percent of the values in the inverted CWM curve were selected, the rest were set to zero. For the red path, a polynomial fit of order O poly,red was fitted to and subtracted from the curve as baseline compensation. All values above T red were selected, the rest were set to zero.
At each position, the largest value from the three curves was selected to create the SDR curve. To avoid the darker streaks in the borders of the images that appear for fully functional transducers, the first and last P exclude percent of the SDR curve were set to zero.
In order to obtain a scalar measure of the nonuniformity of the system, the area under the SDR curve was finally calculated using trapezoidal numerical integration. The area under the curve was normalized by the number of elements in the transducer. An example of an SDR curve is shown in Fig. 4, overlaid on a median image of the 150 images (N stack = 150) selected from Case 1 for creating the SDR curve.
The algorithm was applied to the image data described in  For Case 3 in Fig. 5, there is a pattern in the SDR curves around curves 6300-7400 that cannot be seen in the FirstCall measurement. By visual inspection of the median uniformity image for this period the pattern was found also in the median image, which makes it reasonable to assume that this was an intermittent defect that was present during this period of approximately 1100 images (5 months) for this particular transducer/scanner combination.
The area under the SDR curves for the three cases are shown in However, a limitation of the present study is that this has not been tested.
The algorithm used in the present paper has similarities to the one described in IEC Technical Specification. 24   The number of images used in the image stack for the construction of each SDR curve, N stack , was set to 150 images in the present paper. The higher N stack , the more robust the algorithm becomes for systematic defects that are present in all N stack images, but the longer it takes to replace the images in the stack and thus the longer it may take for a new defect to be detected. N stack was set to 150 in order to get illustrative SDR curves for the cases with few false positives, but this number may be reduced for a good balance between occasional false positives and an early detection of defects.
To detect intermittent defects it is also preferable to have a smaller image stack. However, a single measurement of a phantom or a Using the area under the SDR curve is a simple way to implement an automatic detection of defects. More advanced ways could involve applying weight functions that amplify the central part of the SDR curve more than the peripheral part or that emphasize wider streaks more than narrow ones, etc. In this way, the clinical relevance of the scalar measure might be increased.
F I G . 6. The area under the SDR curve for each SDR curve for the three cases.
The described method has been developed for stored images.
One condition for this approach to be successful is that a given All scanner settings that affect beam forming, image processing, and possible algorithms that the manufacturers have built into the scanner to compensate for defects in the equipment, affect the clinical image and thus the median uniformity image. If spatial compounding is used, this will probably impact performance of the method to be less sensitive in a similar way as spatial compounding impacts visibility of defects when checking horizontal uniformity using a phantom. For the images used in the present paper, there was no information available whether spatial compounding was used or not. Defects that can be detected using clinical images are also present in the final processed clinical images, which is an advantage if the only purpose is to detect aberrations in the images. However, the major advantages of using stored clinical images are that this can be done automatically and that there is no need to get admission to the equipmentonly access to the database where the images are stored is required. If some kind of curves are presented, like the ones in Fig. 5 in the present paper, it is easy to see the location, size and width of the defects and to follow if they are intermittent, steady or increasing.
Assessment of the image uniformity as part of a quality assurance program is used for other imaging modalities as well, such as e.g., computed 26,27 and digital 28 radiography systems and in nuclear medicine. 29,30 To use median images of clinical images for detecting nonuniformities in these systems may also be possible, and could be useful as a complement to quality assurance for early detection of a

| CONCLUSION S
A method of using clinical images for assessment of horizontal uniformity in ultrasound imaging has been introduced. The method is applicable to, e.g., linear array transducers, as used in the present study.
An algorithm for automatic detection of horizontal nonuniformities has been described and tested on more than 31000 clinical images from three systems in a case study. Subjectively, the algorithm seemed to be robust, as visually similar SDR curves were obtained for images acquired at different time points over long periods of time.
Furthermore, the results showed good visual agreement with available electrical measurements of the defective transducers, indicating a potential use of clinical images for early and automatic detection of defective transducers, as a complement to quality assurance.