Automatic measurement of kidney dimensions in two-dimensional ultrasonography is comparable to expert sonographers

Abstract.

Purpose

Length and width measurements of the kidneys aid in the detection and monitoring of structural abnormalities and organ disease. Manual measurement results in intra- and inter-rater variability, is complex and time-consuming, and is fraught with error. We propose an automated approach based on machine learning for quantifying kidney dimensions from two-dimensional (2D) ultrasound images in both native and transplanted kidneys.

Approach

An nnU-net machine learning model was trained on 514 images to segment the kidney capsule in standard longitudinal and transverse views. Two expert sonographers and three medical students manually measured the maximal kidney length and width in 132 ultrasound cines. The segmentation algorithm was then applied to the same cines, region fitting was performed, and the maximum kidney length and width were measured. Additionally, single kidney volume for 16 patients was estimated using either manual or automatic measurements.

Results

The experts resulted in length of $84.8\pm 26.4\mathrm{mm}$ [95% CI: 80.0, 89.6] and a width of $51.8\pm 10.5\mathrm{mm}$ [49.9, 53.7]. The algorithm resulted a length of $86.3\pm 24.4$ [81.5, 91.1] and a width of $47.1\pm 12.8$ [43.6, 50.6]. Experts, novices, and the algorithm did not statistically significant differ from one another ($p>0.05$). Bland–Altman analysis showed the algorithm produced a mean difference of 2.6 mm (SD = 1.2) from experts, compared to novices who had a mean difference of 3.7 mm (SD = 2.9 mm). For volumes, mean absolute difference was 47 mL (31%) consistent with $\sim 1\mathrm{mm}$ error in all three dimensions.

Conclusions

This pilot study demonstrates the feasibility of an automatic tool to measure in vivo kidney biometrics of length, width, and volume from standard 2D ultrasound views with comparable accuracy and reproducibility to expert sonographers. Such a tool may enhance workplace efficiency, assist novices, and aid in tracking disease progression.

1. Introduction

The kidney is an essential organ for life. It performs waste removal, fluid balance, blood pressure regulation, red blood cell production, and calcium homeostasis. In states of disease, its morphology may be altered. For example, in states of chronic injury with significant collagen and fibrosis accumulation, the kidney may contract and shrink in size.^1,2 In states of fluid overload from hydronephrosis or inflammation from pyelonephritis, the kidney may enlarge.^2,3 Changes in the morphological parameters of an individual’s kidney over time may subsequently be used as disease indicators.^2,3 Obtaining these morphological measurements, however, is not as simple as it may appear.

Length and width of the kidney are two often-reported morphological measurements from imaging studies. The normal distribution of these morphologies has been studied by scientists. For example, a study of 300 healthy in vivo kidneys imaged using magnetic resonance imaging (MRI) reported that the kidney lengths in men are $\sim 124\pm 9\mathrm{mm}$ and $115\pm 11\mathrm{mm}$ in women.⁴ The normal range of kidney length is generally considered between 100 and 120 mm, with side-specific differences.^5,6 These biometrics can be used to discriminate between normal and abnormal states. For example, an ultrasound-based measure of kidney length of over 16.5 cm was a prognostic factor in the onset of moderate-to-advanced chronic kidney disease.⁷ Other studies have shown there is a measurable decrease in length between patients with and without chronic kidney disease ($109.4\pm 13.0\mathrm{mm}$ versus $114.4\pm 9.6\mathrm{mm}$).⁸ These ultrasound-based measurements are commonly made using computerized calipers in imaging software. However, the measurement of these dimensions requires a non-trivial amount of expertise. An individual, whether an expert sonographer or novice medical trainee, must recognize the frame with the maximum length of the kidney, then appropriately measure it by identifying the bounds of the capsule and repeat this step for all dimensions. As ultrasound is operator dependent, with potential errors in both image acquisition and interpretation, a degree of uncertainty is introduced in these clinical measurements depending on the individual’s ability to differentiate the capsule from surrounding tissues.

Given the ability to measure length and width of the kidney in different views, single kidney volume (SKV) can be estimated from untracked two-dimensional (2D) ultrasound images alone. Recent literature has established kidney volume as a surrogate biomarker for kidney function, with prognostic implications in disease conditions, such as polycystic kidney disease.^9–11 The normal range for men has been reported as being between 132 to 276 mL and 87 to 223 mL in women.⁴ In diseases, such as autosomal polycystic kidney disease, measurement of a twofold or threefold increase in kidney volume has significant prognostic implication for decline in kidney function.¹⁰ In diseases, such as chronic kidney disease, patients with disease were found to have significantly smaller kidney volumes ($74.7\pm 24.6\mathrm{mL}$) as compared to healthy controls ($171.7\pm 32.6\mathrm{mL}$).¹² Although volumetric imaging modalities, such as computed tomography (CT), MRI, and three-dimensional (3D) ultrasound imaging can provide the 3D information needed to compute volume, these forms of imaging require additional cost, time, may involve ionizing radiation, and may not be suitable for all patients. On the other hand, 2D ultrasound is already the first-line imaging modality for kidney dysfunction given its low-cost, non-ionizing, and real-time nature.¹³ Given that the most common form of kidney ultrasound is freehand two-dimensional imaging, the precise calculation of SKV may necessitate manually annotating individual frames across an entire ultrasound video clip. This process is both time and labor intensive as well as potentially biased to underestimation. Kim et al.¹⁴ demonstrated that, when compared to CT, two-dimensional volumes underestimated the true volume by $\sim 15\%$.

Alternatively, the prolate ellipsoid approach employs length and width in only two imaging views to estimate the overall volume.¹⁵ Specifically, volume is estimated as the product of the length in a longitudinal view, the length in a transverse view, and the width in a transverse view, all multiplied by $\pi /6$. Nevertheless, ultrasound-based calculations of SKV require high-quality imaging acquisition and the ability to manually assess kidney dimensions in conventional views to estimate volume. Studies comparing ultrasound-based kidney volume measures to MRI have revealed a lack of intra- and inter-observer variation.¹⁶ Two-dimensional ultrasonic volume estimates can be difficult to acquire quickly and accurately. For all of these morphological measurements, the ostensibly simple activity may be difficult to accomplish in a reproducible manner and may be time-consuming. Nonetheless, this work is ideally suited for computer-assisted techniques, which can help automate the kidney biometric measuring process utilizing computer vision and machine learning.

With major advancements in semantic segmentation in recent years, it may be possible to automate this task. The U-net architecture introduced by Ronneberger et al.¹⁷ obtained groundbreaking results for medical image segmentation at the time and has subsequently served as the foundation for thousands of newer networks. nnU-net¹⁵ is another notable network in image segmentation.¹⁸ nnU-net in obtained state-of-the-art results on 23 different datasets and tasks spanning many organs and modalities; however, ultrasound was not investigated. There is a substantial amount of literature on segmenting anatomy from ultrasound images, ranging from obstetrics and the pelvic floor,¹⁹ fetal anatomy,^20,21 cardiac anatomy and landmarks,²² and organs such as prostate²³ and liver.²⁴ Specific to kidney ultrasound, Yin et al.²⁵ recently created a fully automatic segmentation neural network for detecting the kidney’s boundary that incorporated learned boundary distance maps to then optimize pixelwise labels. To the best of our knowledge, however, no work has studied the automation of kidney biometric extraction from ultrasound images and compared the results to those of both professionals and trainees, nor evaluated such a system in both native and transplanted kidneys. It is important to consider both native and transplanted kidneys when evaluating a system for measuring kidney biometrics, as there are differences in their anatomical position (retroperitoneum versus the iliac fossa), depth (shallow versus deep), and the imaging frequencies used (higher versus lower). Automating kidney biometrics derived from ultrasound has the potential to reduce variability and make operator experience irrelevant. Furthermore, as image acquisition is commonly performed freehand, this approach avoids potential errors associated with the summation of segmented areas in an ultrasound video, such as overestimation.

The primary contribution of this study is the proof-of-concept and thorough validation of an automatic kidney measuring software tool as outlined in Fig. 1. This software tool is a fully automated algorithm that extracts kidney length, kidney width, and SKV in 2D ultrasound cines of both native and transplanted kidneys. Length and width are examined using 132 cines of renal ultrasonography (16,612 frames total). Sixteen individuals with ultrasonography cines in both standard longitudinal and transverse views were evaluated for volume measures (32 cines, total frames: 4027). Notably, both native and transplanted kidneys are included. Two experienced sonographers and three medical students are compared to the proposed method, and relationships with transducer or patient features are evaluated. Automating the process of evaluating morphology from 2D ultrasound should permit the use of AI-enhanced 2D ultrasound as a reliable and efficient workflow efficiency gain and enable the use of an SKV measurement as a surrogate biomarker.

Fig. 1.

The manual approach of annotating and measuring kidney biometrics in 2D ultrasound compared against the proposed method using a neural network and image processing.

2. Methods and Materials

2.1. Definitions of Kidney Biometrics

In a longitudinal image of the kidney, length $l_L$ is defined as the largest line segment stretching from the capsule at one pole, via the central echogenic complex, to the capsule at the other pole. Similarly, in a longitudinal image, the width $w_L$ defined as the maximum line segment within the kidney capsule that is perpendicular to the length. In a transverse image of the kidney, the length $l_T$ is the greatest line segment stretching from one location on the capsule through the central echogenic complex to another position on the capsule. In the transverse view, the width $w_T$ is defined as in the same manner as $w_L$. These four line segments are pictorially represented in Fig. 2. For SKV, $V$, the ellipsoid method shown to have the best performance by Hammoud et al.¹⁵ was used. This uses three measures of the kidney in longitudinal and transverse views, as defined in the following equation:

$$V=\frac{\pi }{6}*l_L*l_T*w_T.$$ (1)

Fig. 2.

Visual representation of the kidney’s biometrics: (a) a longitudinal view of a kidney with length and width in this view denoted as $l_L$ and $w_L$ and (b) a transverse view of a kidney with length and width denoted as $l_T$ and $w_T$.

2.2. Ultrasound Datasets and Algorithmic Approach

Following permission by the University of British Columbia’s Research Ethics Board (H19-01199), two datasets (Table 1) were produced from images of both native and transplanted kidneys acquired at Vancouver General Hospital (Vancouver, BC, Canada) over a five-year period. The data retrieved consist of 2D kidney ultrasound images in standardized views of longitudinal and transverse orientation. Complete scans in both views may not always be retained due to institutional storage limitations; sometimes only a select few frames are stored. Hence, the ratio of patients to ultrasound cines is not consistently 1:2. There are two datasets for evaluation: one for kidney lengths and widths and one for SKVs. The first dataset includes an additional 132 ultrasound cines from 92 different subjects, and measurements of length and width are considered distinct for patients in different views. The second dataset includes 32 ultrasound images from 16 subjects, and each patient has two ultrasound cines, one for each of longitudinal and transverse views.²⁶ There was no overlap of patients in the “lengths and widths” and “volumes” datasets.

Table 1.

Dataset descriptions for evaluation of the automatic methods.

	# Subjects	Native kidney	Transplant kidney	# Cines	Longitudinal view	Transverse view
Lengths and widths	92	61	31	132	66	56
Volumes	16	11	5	32	16	16

First, an nnU-net segmentation neural network¹⁸ was trained using a publicly available ultrasound kidney dataset.²⁶ The data used to train this network consists of 514 2D kidney ultrasound images from the same number of patients in standardized views of longitudinal and transverse. This dataset and its detailed description, along with segmentation network validation, are publicly available.²⁶ Two expert sonographers with $>20$ years of experience in abdominal ultrasonography independently and manually annotated the entire kidney capsule in each of these images using the VGG Image Annotator (Oxford, UK).²⁷ A nnU-net model was trained using an 80:20 train–test ratio to segment the kidney’s capsule in 2D ultrasound images.¹⁸ The ground truth annotation was randomly selected between the two sets of annotations with equal probability of being selected. This algorithm by Isensee et al.¹⁸ uses a combination of dataset fingerprinting and heuristics to adapt to the provided training data. Validation was performed using fivefold cross validation and an ensemble method for inference. University of British Columbia’s Advanced Research Computing Sockeye platform was used to perform all training, using a single 32 GB NVIDIA Tesla V100 GPU.

Second, on the 132 ultrasound cines from 92 patients, the two expert sonographers and three medical trainees with $<2$ years of kidney ultrasound interpretation experience, performed manual measurements of kidney length and width. For each cine, the annotator independently reviewed the entire cine and selected the frame that they visually ascertained contained the maximum kidney length. They then annotated length and widths in the selected frame using the VGG Image Annotator. The definitions of length and width in longitudinal and transverse view were provided to the annotator. In comparison, the trained nnU-net model was applied to each cine in its entirety. A binary mask of the kidney capsule in each frame was generated. The frame with the largest mask size was automatically selected from the cine and a bounding rectangle fit onto it. The rectangle’s major axis and minor axis were treated as the kidney length and width, respectively, for that cine and thus patient. The comparison of the manual workflow versus the automated workflow is depicted in Fig. 1.

Complete full scans in both views may not always be retained due to institutional storage limitations. Thus to evaluate volume estimations, a different group of 16 patients having longitudinal and transverse video clips acquired in the same imaging session were chosen (32 clips). The prolate ellipsoid approach was utilized to calculate volume using both manual and automatic measurements obtained by the segmentation network.¹⁵

To assess for potential confounding variables, imaging variables including transducer frequency, imaging depth, and ultrasound manufacturer and model were extracted from the retrieved original DICOM data. Similarly, patient variables, such as age, sex, estimated glomerular filtration rate (eGFR), body mass index (BMI), and primary diagnosis were collected using electronic health records data that was cross-referenced to the imaging files.

2.3. Data Analysis

First, the intra- and inter-rater variability of the expert and trainee groups was evaluated using the 132 cines for length and width. Randomly selecting 20 images, the intra-rater variability for each dimension was calculated. Similarly, the inter-rater variability of each annotator group was compared.

Our primary hypothesis is that the automatic method is not significantly different than the expert measurements within a bias of 5 mm. This value is based on the natural variation of kidney lengths being reported as on the order of 10 mm.²⁸ Our secondary hypothesis is that novices have a significantly larger mean measurement difference from experts when comparing against the mean difference of the automatic method to experts. Using D’Agostino-test Pearson’s with a significance threshold of 0.05, the normality was determined. For normal distributions, the Student’s paired $t$-test was used to evaluate statistically significant differences, whereas the Mann–Whitney non-parametric test was employed for non-normal distributions. All morphology measurements are provided as means standard deviations with confidence intervals of 95% for each group, and then statistically significant differences are analyzed. Furthermore, Bland–Altman analysis was performed in experts versus the automatic method to evaluate for systematic measurement error. Absolute and relative differences in the volume measurements are also presented.

Finally, Pearson’s correlation coefficient and analysis of variance were used to evaluate potential confounding factors and their effects on the obtained measurements related to image acquisition (frequency, depth, manufacturer, and machine type) as well as patient characteristics (age, sex, BMI, eGFR, and primary diagnosis).

3. Results

3.1. Segmentation Accuracy and Inter-Rater Variability

The Dice–Sorenson coefficient (DSC) of the neural network in segmenting the kidney capsule was $0.86\pm 0.01$. The inter-rater variability of the two experts was 1.5 mm, whereas the inter-rater variability of the three novices was 6.3 mm for both lengths and widths. Intra-rater variability was 1.0 and 6.6 mm for the experts and novice groups, respectively.

3.2. Measurements and Differences Between Groups

The experts resulted in a mean length of $84.8\pm 26.4\mathrm{mm}$ [95% CI: 80.0, 89.6] and a mean width of $51.8\pm 10.5\mathrm{mm}$ [95% CI: 49.9, 53.7] across all images. Comparatively, the novices resulted in a mean length of $82.3\pm 22.9\mathrm{mm}$ [95% CI: 80.0, 84.9] and a mean width of $46.5\pm 11.3\mathrm{mm}$ [95% CI: 45.2, 47.8]. There was no statistically significant difference between the novices and the experts in paired analysis ($p>0.05$). The algorithm resulted in a length of $86.3\pm 24.4\mathrm{mm}$ [95% CI: 81.5, 91.1] and a width of $47.1\pm 12.8\mathrm{mm}$ [43.6, 50.6]. There was no statistically significant difference between the automated algorithm and the experts in pairwise analysis ($p>0.05$). Bland–Altman analysis demonstrated a mean difference of 2.6 mm (SD = 1.2 mm) in experts versus the algorithm. The 95% limits of agreement ranged from 0.17 to 5.03 mm. In contrast, when comparing experts versus novices, analysis demonstrated a mean difference of 3.7 mm (SD = 2.9 mm) with limits of agreement between $-1.98$ to 9.38 mm.

For SKV, mean absolute difference of estimates using automated measurements to estimates using expert measurements was 47 mL while the mean relative difference was 31%.

3.3. Correlations with Transducer and Patient Factors

Only imaging depth had statistically significant relationships for one type of measurement in only one group—the kidney width as estimated by novices—among the image acquisition factors investigated. The sole significant correlation was between imaging depth and novice-measured kidney width was $r=0.27$ ($p<0.05$). There was no significant correlation found between imaging depth or imaging frequency and the mean differences. There was no association between any other group’s measurement and imaging frequency, manufacturer, or machine type. In addition, when patient characteristics such as age, sex, primary diagnosis, eGFR, and BMI were evaluated, no statistically significant correlations were found between any parameter with any group.

4. Discussion

Our proposed approach for the automatic extraction of kidney biometry from 2D ultrasound images using a segmentation neural network achieves a measurement accuracy comparable to experts. In particular, the pairwise mean difference between measurements of the two methods was 2.6 mm, close to the inter-rater variability of the experts of 1.5 mm. The automatic extraction method significantly outperforms the three novices who achieved a mean difference of 3.7 mm and had a high inter-rater variability of 6.3 mm. The automatic method’s mean difference is comparable to previously reported intra-observer variation with 2.09 mm for the right kidney and 0.89 mm for the left kidney.⁸ There was no discernible difference in performance for length versus width or between the type of standardized view. It is feasible that a further improvement in DSC of the segmentation network may yield improved measurements, as the segmentation network demonstrated a tendency to over-segment the kidney capsule. This results in an overestimate for kidney length, supported by the limits of agreement being between 0.17 and 5.03 mm. A detailed tolerance analysis and comparison with other existing segmentation models would assist in understanding the relationship between segmentation accuracy and measurement accuracy.

In evaluating SKV, while on a small set of patients, the automatic approach with the prolate ellipsoid method performed well achieving a 47 mL absolute difference. Although the mean relative difference is moderate at 31%, this is comparable to relative errors previously reported in 2D ultrasonic volume calculations of the kidney, such as the 24% reported by Bakker et al.²⁹ However, this error is still large. Differences in volume change, such as the 19.5% change over one year that has been reported in kidney donors post-nephrectomy,³⁰ may be missed using this proof-of-concept approach. It is important to recognize that small errors in any one dimension quickly yields larger errors in volume calculation even though volume is linear with respect to each dimension. For example, following the prolate ellipsoid method, a 1 mm error in one dimension yields a relative difference of 10%. Similarly, either a 3 mm error in one dimension or a 1 mm error in all three dimensions yield a relative difference of 30%. Furthermore, it is well understood that the kidney is not an ellipsoid in reality, and there is a degree of error associated with this simplification. Nonetheless, given the modest sample size, it is worthwhile to investigate the 2D ultrasound-based SKV measurement prospectively and longitudinally. Particularly, frequent measurements of the same patient may highlight ultrasound’s ability to evaluate kidney parameters consistently throughout time, hence diminishing the impact of the error between sonographers and the algorithmic technique. This technique overall makes minimal assumptions and requires minimal processing.

Surprisingly, the transducer factors and patient factors had minimal impact on measurements. For example, one may expect that with lower frequencies or greater imaging depths, the decrease in ultrasound spatial resolution may impact the ability of a human or machine to delineate the kidney capsule. Similarly, in patients with higher BMI, the ultrasound image quality may be poorer and similarly impact the segmentation or measurements. In the small datasets used in this study, neither seems to have a significant influence on kidney biometric measurements. Furthermore, there were no differences in performance between native or transplanted kidney images. As the transplanted kidney is in a different anatomical location (the iliac fossa), and thus shallower relative to the skin surface and higher imaging frequencies, it is again surprising to observe no substantial differences. Although no significant factors were found, future work could consider the use of multiple regression to assess the contribution of various factors on the measurements.

An overarching issue for this study however is the lack of a gold standard for the measurements, such as MRI or CT. Although our data did not have paired ultrasound-MRI/CT for each patient, the use of a volumetric imaging modality would provide a reference by which comparisons between the expert, novices, and algorithm could be made. As presented, the human expert is treated as the reference standard. Furthermore, previous literature has demonstrated that ultrasonography can underestimate the true volume of an organ when compared to MRI/CT.^14,29

While not formally evaluated, the time for a human to perform cine review and annotation can be estimated to be 1 to 5 min. In contrast, inference time for a single frame is 10 ms, corresponding to an average of 1.2 s for an entire cine; a substantial improvement to achieve comparable results.

5. Conclusions

We proposed an automatic tool to measure kidney biometry from standard ultrasound views in both native and transplanted kidneys. This was achieved using a segmentation neural network for extracting the kidney capsule, followed by region fitting where major and minor axes of a bounding rectangle modeled kidney length and width. The proposed tool was comparable to experts in measurement accuracy with a mean difference of 2.6 mm and achieved less bias and variance than novice measurements. Future work includes evaluating automatic biometric extraction techniques for non-standard views, exploring complete end-to-end automatic measurements without an intermediate segmentation step, and prospective studies to understand clinical performance and translatability. Such a tool may enhance workplace efficiency, assist novices, and be useful in tracking disease progression.

Biographies

Rohit Singla received his BASc degree in computer engineering and his MASc degree in biomedical engineering from the University of British Columbia (UBC) in 2015 and 2017, respectively. He is an MD/PhD student in biomedical engineering at the same university. He is the recipient of the Vanier Graduate Scholarship and the Kidney Foundation of Canada Fellowship. He is the author of more than 20 peer-reviewed articles. His current research interests include ultrasound image analysis, non-invasive tissue characterization, and kidney disease.

Cailin Ringstrom received her bachelor of applied science in engineering physics from UBC in 2022. She is a master’s student in robotics, systems, and control at ETH Zurich. Her current interests include medical robotics and computer vision.

Ricky Hu received his BASc degree in engineering physics and his MASc degree in biomedical engineering from UBC in 2016 and 2019, respectively. He is a medical student at Queen’s University. He is an author of more than 20 publications in the fields of artificial intelligence in texture analysis, physics in medical imaging, medical robotics, and medical education.

Zoe Hu received his MD from Queen’s University and is a radiology resident at the same institution.

Victoria Lessoway has more than 45 years of ultrasound experience working as a sonographer. She has co-authored more than 50 publications, including the development of fetal biometry tables that were a provincial standard for more than 40 years. She is semi-retired and is currently the co-executive director of the BC Ultrasonographers’ Society. In 2021, she retired in good standing from the American Registry of Diagnostic Medical Sonographers and is currently a certified member of Sonography Canada.

Janice Reid: Biography is not available.

Christopher Nguan is a kidney transplant urologist, operating at Vancouver General Hospital since 2006, where he is the surgical director of renal transplantation. He is a surgeon-scientist, appointed as an associate professor in the Department of Urologic Sciences at the UBC. His main areas of research are twofold: research in applied sciences revolving around advanced technologies for medicine and surgery including computer vision, AR/VR, diagnostic imaging, and machine learning, and a second research arm in basic and translational science related to cell/tissue/organ preservation, diagnostics, organ modification, and transplant outcomes. He has authored more than 90 peer-reviewed publications and is named on six novel patents.

Robert Rohling received his BASc degree in engineering physics from UBC, his MEng degree in biomedical engineering from McGill University, and his PhD in information engineering from the University of Cambridge. He is a professor in the Department of Electrical and Computer Engineering and the Department of Mechanical Engineering at UBC. He is also the director of the Institute for Computing, Information, and Cognitive Systems. His research is in ultrasound including elastography and microfabrication.

Disclosures

The authors declare no conflicts of interest.

Code, Data, and Materials Availability

Due to the sensitive nature of the medical imaging data used in this study, data are unavailable.