AI IN BRIEF

Deployed Deep Learning Kidney Segmentation for

Polycystic Kidney Disease MRI
Akshay Goel, MD • George Shih, MD • Sadjad Riyahi, MD • Sunil Jeph, MD • Hreedi Dev • Rejoice Hu • 
Dominick Romano, BS  •  Kurt Teichman, BS, MEng  •  Jon D. Blumenfeld, MD  •  Irina Barash, MD, MS  • 
Ines Chicos, MS CCRC  •  Hanna Rennert, PhD  •  Martin R. Prince, MD, PhD
From the Departments of Radiology (A.G., G.S., S.R., S.J., H.D., R.H., D.R., K.T., M.R.P.), Internal Medicine (J.D.B., I.B., I.C.), and Pathology and Laboratory Medicine
(H.R.), Weill Cornell Medicine, 525 E 68th St, New York, NY 10021. Received July 21, 2021; revision requested September 21; revision received January 20, 2022; accepted
January 25. Address correspondence to A.G. (e-mail: akg9006@med.cornell.edu).
Supported by the Weill Cornell Medicine Clinical & Translational Science Center and the Shaw Foundation.
Conflicts of interest are listed at the end of this article.

Radiology: Artificial Intelligence 2022; 4(2):e210205 • https://doi.org/10.1148/ryai.210205 • Content codes:

This study develops, validates, and deploys deep learning for automated total kidney volume (TKV) measurement (a marker of disease severity)
on T2-weighted MRI studies of autosomal dominant polycystic kidney disease (ADPKD). The model was based on the U-Net architecture
with an EfficientNet encoder, developed using 213 abdominal MRI studies in 129 patients with ADPKD. Patients were randomly divided into
70% training, 15% validation, and 15% test sets for model development. Model performance was assessed using Dice similarity coefficient
(DSC) and Bland-Altman analysis. External validation in 20 patients from outside institutions demonstrated a DSC of 0.98 (IQR, 0.97–0.99)
and a Bland-Altman difference of 2.6% (95% CI: 1.0%, 4.1%). Prospective validation in 53 patients demonstrated a DSC of 0.97 (IQR,
0.94–0.98) and a Bland-Altman difference of 3.6% (95% CI: 2.0%, 5.2%). Last, the efficiency of model-assisted annotation was evaluated on
the first 50% of prospective cases (n = 28), with a 51% mean reduction in contouring time (P , .001), from 1724 seconds (95% CI: 1373,
2075) to 723 seconds (95% CI: 555, 892). In conclusion, our deployed artificial intelligence pipeline accurately performs automated segmenta-
tion for TKV estimation of polycystic kidneys and reduces expert contouring time.

ClinicalTrials.gov identification no.: NCT00792155

Supplemental material is available for this article.

© RSNA, 2022

Material and Methods

A utosomal dominant polycystic kidney disease (AD-
PKD) is the most common inherited renal disease,
affecting 12.5 million people worldwide and compris- Patient Datasets
ing 5%–10% of end-stage renal disease (1). Total kid- This Health Insurance Portability and Accountability Act–
ney volume (TKV) indexed to patient height (htTKV) compliant study was approved by the institutional review
(2,3) is an important imaging biomarker for assessing board. Individuals enrolled in the Rogosin Institute AD-
ADPKD severity (4,5). The Mayo classification prog- PKD Repository study provided signed informed consent.
nosticates the time to requiring dialysis for patients Existing axial MRI studies (n = 213 in 129 patients) were
with ADPKD with typical diffuse cystic kidney disease utilized for model development and retrospective valida-
(Mayo class A) using htTKV at a single time point, ad- tion. MRI studies in another 20 patients with ADPKD,
justed for age and estimated kidney growth rate (6,7). performed at other institutions, were utilized for exter-
Tolvaptan can slow the rate of progression of chronic nal validation. MRI studies performed beginning June
kidney disease in patients at risk for rapid progression 1, 2021, (n = 53) were included in a prospective assess-
on the basis of age, creatinine level, and htTKV (8). ment of the clinical implementation. This study was reg-
The importance of htTKV for determining tolvaptan istered on ClinicalTrials.gov with identification number
therapy eligibility and for following patients taking NCT00792155.
tolvaptan has resulted in increased demand for MRI
measurements of TKV. As TKV increases only 2%–5% MRI Protocol
per year (9), a highly reproducible, operator-indepen- All participants used for model training and prospective
dent method for measuring kidney volumes is needed. evaluation underwent a standardized imaging protocol
In this study, we developed, validated, and deployed at 1.5 T using a body phased‐array coil, including axial
a segmentation model using a U-Net architecture with T2‐weighted single-shot fast spin-echo and steady-state
an EfficientNet encoder for accurate polycystic kidney free precession (SSFP) images (details described in Ap-
segmentation at MRI. Furthermore, we evaluated the pendix E1 [supplement]). Participants used for external
efficiency improvement of model-assisted annotation. validation all underwent axial T2-weighted imaging with

This copy is for personal use only. To order printed copies, contact reprints@rsna.org
Deployed Deep Learning Kidney Segmentation for Polycystic Kidney Disease MRI

accuracy at the polycystic kidney boundary. Annotation voxels

Abbreviations were summed to derive TKV.
ADPKD = autosomal dominant polycystic kidney disease, DICOM
= Digital Imaging and Communications in Medicine, DSC = Dice
similarity coefficient, htTKV = TKV indexed to patient height, Data Stratification
NIfTI = Neuroimaging Informatics Technology Initiative, PACS = Patients (n = 129 with 213 studies) were split on the basis of
picture archiving and communication system, SSFP = steady-state patient ID for model development into 70% training, 15%
free precession, TKV = total kidney volume
validation, and 15% holdout test sets, with a stratified ran-
Summary dom split controlling for the statistical distribution of TKV.
A deep learning model was developed and deployed to automatically The training, validation, and test sets contained 154 MRI stud-
segment kidneys in autosomal dominant polycystic kidney disease ies (5570 DICOM images), 28 MRI studies (1312 DICOM
MRI studies for determining total kidney volume, a metric of disease
severity.
images), and 31 MRI studies (1050 DICOM images), respec-
tively. No patient was represented in more than one split. The
Key Points dataset split was fixed prior to model training to prevent data
n Model performance on validation MRI studies (53 prospective, leakage. Cross-validation (k-fold) was not performed because
20 external institution) demonstrated a Dice similarity coefficient of limited computing resources and our inclusion of prospec-
greater than 0.97 (quartile Q1  0.94) and Bland-Altman mean tive and external validation.
percentage difference in total kidney volume (model vs manual
reference) less than 3.6% (95% CI: 2.0%, 5.2%).
n Model deployment into the clinical pipeline was accomplished by Model Development, Data Augmentation, and Training
pushing new MRI studies to a clinical server within the picture Procedure
archiving and communication system firewall that runs inference
MRI studies and ADPKD kidney segmentations were used to
on axial T2-weighted sequences and outputs results formatted for
radiologist label refinement and final reporting. train models for the objective of kidney segmentation (Fig 1). AD-
n Prospective assessment in 53 patients showed a 51% reduction in PKD kidney segmentation was formulated as a two-dimensional
radiologist time for model-assisted segmentation compared with problem, such that each slice in an MRI study provided a single
segmenting without the model (P , .001). sample. To generate a three-dimensional segmentation, inference
Keywords was performed on all slices in a volume and then combined. The
Convolutional Neural Network (CNN), Segmentation, Kidney model architecture was based on the encoder-decoder U-Net (14),
with EfficientNet comprising the encoder component (15); Ef-
ficientNet has shown top performance on various image segmen-
various MRI machines (Table E1 [supplement]). tation tasks (16). Further details on the model architecture are
shown in Table E2 and Figure E1 (supplement).
Code, Model, and Data Availability Training data were augmented using affine transformations
All code was developed in the Python programming language (17). Model checkpoints were saved on the basis of improving
(version 3.8.5) using the PyTorch deep learning framework performance on the validation set that was based on the met-
(version 1.7.1) (10). The training code, inference code, and ric, predicted-TKV error. Upon the completion of training, the
model weights for this project are publicly available on our single best model checkpoint was evaluated against the internal
code repository (https://github.com/aksg87/adpkd-segmenta- holdout test dataset. Further details on model development, data
tion-pytorch) (11). Data availability is discussed at the end of augmentation, and training procedure are discussed in Appendix
the article. E1 (supplement).

Data Preparation Model Deployment into the Clinical Routine and Prospective
De-identified Digital Imaging and Communications in Medi- Validation
cine (DICOM) axial T2-weighted and SSFP MRI studies were The model was deployed on a clinical Linux server within the
converted into NIfTI (Neuroimaging Informatics Technology picture archiving and communication system (PACS) firewall
Initiative) file format using the NiBabel package in Python (Fig 1). Remote connection to the server for the purposes of
(https://nipy.org/nibabel/) (12). Segmentation masks were saved annotation and validation was achieved using XFCE (18),
in NIfTI format within ITK-SNAP. Normalization was ap- an open-source desktop environment. The single best model
plied to each scan by linearly scaling 16-bit grayscale values checkpoint was utilized after training and fixed throughout ex-
to [0, 1] (single-precision floating point), on the basis of the ternal and prospective validation. To summarize the data flow,
minimum and maximum grayscale value. first, new MRI studies were pushed to the Linux server by a
radiologist via our clinical PACS. Then, inference was triggered
TKV Reference Standard upon the arrival of new axial T2-weighted images. Next, model
The reference standard TKV was obtained using ITK-SNAP predictions were saved as NIfTI files, fully editable (using ITK-
(13) to manually segment kidneys by a radiology fellow (A.G.) SNAP 3.8.0 running on the Linux server) by a radiologist to
and radiology attending physician (M.R.P.) with a combined correct any errors, and time savings of model-assisted annota-
25 years of experience imaging ADPKD. All annotations were tion versus manual annotation were recorded for the first con-
performed without interpolation with attention to detail for secutive 50% of prospective cases. The possibility of bias from

2radiology-ai.rsna.org  n  Radiology: Artificial Intelligence Volume 4: Number 2—2022

 Goel et al

Figure 1:  Schematic summarizes project infrastructure on deep learning server. Training is highlighted with a light pink background. Deployment and inference are high-
lighted with a light blue background.

repeated annotation of the same scan was minimized by hav- internal validation. Model performance was validated on ex-
ing a minimum 1-week interval between manual and model- ternal data with a median DSC of 0.98 (IQR, 0.97–0.99) and
assisted annotation. mean percentage difference in TKV of 2.6% (95% CI: 1.0%,
4.1%). Additional validation was performed on prospective in-
Statistical Analysis ternal data with a median DSC of 0.97 (IQR, 0.94–0.98) and
Comparison between the manual reference standard (ground mean percentage difference in TKV of 3.6% (95% CI: 2.0%,
truth) and the model was performed using Dice similarity coef- 5.2%) (Fig 2, Table 2). Model performance on the internal test
ficient (DSC) defined as set is shown in Figure E3 (supplement).
Multiple visualizations were evaluated to assess performance.
Model predictions and the reference standard were visualized
, with three-dimensional overlapping surface renderings (Fig 3).
Predictions were also visualized as two-dimensional images (Figs
E2, E4 [supplement]).
where A and B represent the ground truth and model-derived Time savings analysis in prospective patients revealed that
segmentations, respectively (19). Mean absolute error and manual contouring (starting with no segmentations) required
Bland-Altman analysis were used to assess agreement in TKV a mean of 1724 seconds (95% CI: 1373, 2075), compared
calculation. The significance of time savings of model-assisted with 723 seconds (95% CI: 555, 892) for model-assisted
annotation was assessed with the (paired) Student t test. contouring, a 51% reduction in time for model-assisted con-
touring (P , .001).
Results Failure analysis was conducted to evaluate the largest model
Table 1 shows baseline characteristics of the patients with errors. Examples of the most significant prospective errors and
ADPKD used for model development (n = 129), outside the corresponding radiologist corrections are shown in Figure
external test cohort (n = 20), and internal prospective test 4, which include a fluid-filled stomach partially labeled as left
cohort (n = 53). cystic kidney, the urinary bladder labeled as cystic kidney, and
Figure 2 shows Bland-Altman plots for agreement in TKV mischaracterization of cysts at the border of liver and right kid-
calculations and DSC by TKV plots for external and prospective ney, among other examples.

Radiology: Artificial Intelligence Volume 4: Number 2—2022  n  radiology-ai.rsna.org 3

Deployed Deep Learning Kidney Segmentation for Polycystic Kidney Disease MRI

Table 1: Characteristics of Development and Validation Data

Cohort Development Sets* External Validation Prospective Validation

No. of patients 129 20 53
No. of DICOM files 7932 894 3893
Age at scan (y)† 47 6 13 48 6 13 47 6 16
Sex
 Female 65 (50) 10 (50) 29 (55)
 Male 64 (50) 10 (50) 24 (45)
eGFR (mL/min/1.73 m2)† 72.2 6 30 59 6 20 72 6 30
htTKV at scan (mL/m)‡ 787 (425–1324) 759 (578–1535) 563 (164–4492)
Genotype
 PKD1 80 (62) NC NC
  Nontruncating 30 (23)
  Truncating 50 (39)
 PKD2 28 (22) NC NC
  No mutation detected 21 (16) NC NC
Mayo class
 A 14 (11) 4 (20) 9 (17)
 B 35 (27) 4 (20) 17 (32)
 C 44 (34) 4 (20) 15 (28)
 D 24 (19) 3 (15) 8 (15)
 E 12 (9) 5 (25) 4 (8)
Race
 Asian 3 (2) 2 (10) 2 (4)
 Black 5 (4) 1 (5) 1 (2)
 White 115 (89) 16 (80) 34 (64)
 Unknown 6 (5) 1 (5) 16 (30)
Note.—Except where otherwise noted, data are numbers with percentages in parentheses. DICOM = Digital Im-
aging and Communications in Medicine, eGFR = estimated glomerular filtration rate, htTKV = height-adjusted
total kidney volume, NC = not collected.
* Development sets include the training, validation, and hold-out test sets used for model development.
†
Data are mean 6 SD.
‡
Data are median with IQR in parentheses.

Discussion only three orthogonal measurements but is far less accurate and
Imaging patients with ADPKD demands accurate, reproduc- less reproducible (7,19). Operator exhaustion from the tedious
ible kidney volume measurements for calculating htTKV. These contouring task may naturally lead to human error, particularly
data from 173 patients with ADPKD (129 development, 20 given the precision required at contour borders. Alternatively, in
external validation, and 53 acquired after deployment) dem- model-assisted annotation as deployed in this study, we modify
onstrate a deep learning segmentation model with exceptional the contouring task significantly and reduce operator fatigue. We
multi-institutional and postdeployment prospective perfor- believe this may enable radiologists to produce more accurate
mance. The prospective and external validation demonstrated contours of kidneys in ADPKD.
excellent DSC greater than 0.97 and a low Bland-Altman As established in many studies, deep learning excels at com-
mean percentage difference in TKV less than 3.6%. Moreover, puter vision tasks (20–22), but deep learning studies often
model-assisted segmentation required 51% less time compared lack prospective validation and often fail to impact radiologic
with manual contouring without model assistance. clinical routines. In contrast, our study includes external and
Conventionally, planimetry tracing (2) and stereologic meth- postdeployment prospective validation and incorporates rec-
ods (22) are used to measure TKV at MRI. These methods ommendations from the Checklist for AI in Medical Imaging
comprise a two-step process in which an initial organ contour (or, CLAIM) guidelines (23). While our model architecture
is drawn, followed by tedious slice-by-slice correction of contour is novel, by applying an EfficientNet encoder to ADPKD kid-
errors. Additionally, the manual ellipsoid method of calculat- ney segmentation, our study’s primary contribution is dem-
ing TKV has attempted to simplify kidney measurement using onstrating clinical impact from model-assisted annotation in

4radiology-ai.rsna.org  n  Radiology: Artificial Intelligence Volume 4: Number 2—2022

 Goel et al

Figure 2:  External dataset (top) and prospective dataset (bottom) validation with Bland-Altman agreement analysis and Dice
similarity coefficient by htTKV. BA = Bland-Altman, htTKV = TKV indexed to patient height, TKV = total kidney volume.

Table 2: Internal Validation Set and Test Set Results and Dataset Characteristics

Development: Validation:
Parameter Validation Set Development: Test Set Validation: External Set Prospective Set
Measured TKV (cc) 1586 (843–2615) 1565 (671–2324) 1270 (764–2257) 973 (553–2134)
Predicted TKV (cc) 1539 (826–2557) 1584 (704–2334) 1197 (781–2112) 969 (523–2097)
Dice score 0.95 (0.94–0.96) 0.95 (0.94–0.96) 0.98 (0.97–0.99) 0.97 (0.94–0.98)
Mean difference (cc)* 20 (4, 36) 18 (–2, 38) 29 (2, 57) 72 (28, 114)
Mean % difference* 0.8% (–0.5%, 2.2%) 0.5% (–0.7%, 1.6%) 2.6% (1.0%, 4.1%) 3.6% (2.0%, 5.2%)
Note.—Except where otherwise noted, data are median with IQR in parentheses. TKV = total kidney volume.
* Data in parentheses are 95% CIs.

a real-world health care setting. Prior studies, such as those While our model had a robust performance, our failure anal-
of Kline et al (24) and Sharma et al (25), have published ysis demonstrated notable errors during select clinical scenarios
impressive results on similar tasks; however, they do not show (ie, fluid-filled stomach, distended urinary bladder, hemorrhagic
our external multi-institute validation and time savings when renal cysts, and hepatic cysts at the border of liver and right kid-
deployed into the clinical routine. Furthermore, our code re- ney). Despite these edge cases, our model-assisted annotation
pository, model weights, and select experiments are available saves substantial radiologist time, as at baseline, the radiologist
for other researchers (11). performs manual contouring at our institution. Furthermore,

Radiology: Artificial Intelligence Volume 4: Number 2—2022  n  radiology-ai.rsna.org 5

Deployed Deep Learning Kidney Segmentation for Polycystic Kidney Disease MRI

Figure 3:  Prospective dataset (top row) and external dataset (bottom row) example surface renderings of ground truth reference (red) and
model prediction (blue) volumes with 50% opacity. Overlapping concordant predictions are visualized in shades of purple. Yellow mannequin illus-
trates orientation of the surface renderings.

Figure 4:  Examples of the most significant prospective model inference errors and the corresponding radiologist corrections. Inference label is red, radiologist additions
are green, and radiologist subtractions are indicated by blue arrows. (A) Fluid-filled stomach partially labeled as left cystic kidney. (B) Urinary bladder labeled as cystic
kidney. (C) Liver cyst labeled as kidney. (D) Renal cyst at liver border missed by inference. (E) Complex hemorrhagic left renal cyst incompletely labeled. (F) Collapsed
descending colon labeled as left kidney. (G) Renal cyst at liver border missed by inference. (H) Left elbow medial epicondyle fat labeled as left kidney in a patient imaged
with arms in the field of view.

6radiology-ai.rsna.org  n  Radiology: Artificial Intelligence Volume 4: Number 2—2022

 Goel et al

model performance is likely to improve with retraining our 8. Torres VE, Chapman AB, Devuyst O, et al. Tolvaptan in patients with
autosomal dominant polycystic kidney disease. N Engl J Med 2012;367(25):
model on radiologist-corrected annotations and via ensemble 2407–2418.
learning methods. Last, our study evaluates model performance 9. Grantham JJ, Chapman AB, Torres VE. Volume progression in autosomal
at a single time point, and ultimately assessing kidney growth dominant polycystic kidney disease: the major factor determining clinical
outcomes. Clin J Am Soc Nephrol 2006;1(1):148–157.
over time remains an important clinical end point. We plan to 10. Paszke A, Gross S, Massa F, et al. PyTorch: An Imperative Style, High-
address this limitation with direct assessment of automated AD- Performance Deep Learning Library. arXiv:1912.01703 [preprint] http://
PKD kidney growth tracking in the near future. arxiv.org/abs/1912.01703. Posted December 3, 2019. Accessed April 14,
2021.
11. Goel A. ADPKD Segmentation in PyTorch. https://github.com/aksg87/
Data sharing: Data generated or analyzed during the study are available from the
adpkd-segmentation-pytorch/tree/v1.0. Published 2021. Accessed May 2021.
corresponding author by request and institutional review.
12. Brett M, Hanke M, Cipollini B, et al. nibabel: 2.1.0. Zenodo. https://zenodo.
org/record/60808. Published August 24, 2016. Accessed May 2021.
Author contributions: Guarantors of integrity of entire study, A.G., M.R.P.; study 13. Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour
concepts/study design or data acquisition or data analysis/interpretation, all authors; segmentation of anatomical structures: significantly improved efficiency and
manuscript drafting or manuscript revision for important intellectual content, all reliability. Neuroimage 2006;31(3):1116–1128.
authors; approval of final version of submitted manuscript, all authors; agrees to 14. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for
ensure any questions related to the work are appropriately resolved, all authors; lit- Biomedical Image Segmentation. arXiv:1505.04597 [preprint] http://arxiv.
erature research, A.G., S.R., S.J., M.R.P.; clinical studies, R.H., J.D.B., I.B., I.C., org/abs/1505.04597. Posted May 18, 2015. Accessed April 14, 2021.
H.R., M.R.P.; statistical analysis, A.G., G.S., S.R., S.J., H.D., R.H., M.R.P.; and 15. Tan M, Le QV. EfficientNet: Rethinking Model Scaling for Convolu-
manuscript editing, A.G., G.S., S.R., S.J., R.H., D.R., J.D.B., I.B., H.R., M.R.P. tional Neural Networks. arXiv:1905.11946 [preprint] http://arxiv.org/
abs/1905.11946. Posted September 11, 2020. Accessed November 18, 2020.
Disclosures of conflicts of interest: A.G. No relevant relationships. G.S. Lead- 16. Siddique N, Paheding S, Alom MZ, Devabhaktuni VK. Recurrent residual
ership roles as co-chair of Society for Imaging Informatics in Medicine machine U-Net with EfficientNet encoder for medical image segmentation. In:
learning committee, co-chair for Society of Abdominal Radiology AI committee, Alam MS, ed. Proceedings of SPIE: SPIE defense 1 commercial sensing
co-director of Radiological Society of North America AI certificate course; assis- 2021—pattern recognition and tracking XXXII. Vol 11735. Bellingham,
tant editor of Radiology: Artificial Intelligence. S.R. No relevant relationships. S.J. Wash: International Society for Optics and Photonics, 2021; 117350L.
No relevant relationships. H.D. No relevant relationships. R.H. No relevant rela- 17. Buslaev A, Iglovikov VI, Khvedchenya E, Parinov A, Druzhinin M, Kalinin
tionships. D.R. No relevant relationships. K.T. No relevant relationships. J.D.B. AA. Albumentations: fast and flexible image augmentations. Information
Grants or contracts from Vertex Pharmaceuticals; secretary of American Journal of (Basel) 2020;11(2):125.
Hypertension. I.B. No relevant relationships. I.C. No relevant relationships. H.R. 18. Fourdan O. Xfce: A Lightweight Desktop Environment. In: Proceedings
No relevant relationships. M.R.P. No relevant relationships. of the 4th Annual Linux Showcase & Conference, Atlanta, Atlanta, GA,
October 10–14, 2000. Berkeley, Calif: USENIX Association, 2000. https://
www.usenix.org/conference/als-2000/xfce-lightweight-desktop-environment.
References 19. Higashihara E, Nutahara K, Okegawa T, et al. Kidney volume estimations with
1. Chapman AB, Devuyst O, Eckardt KU, et al. Autosomal-dominant poly- ellipsoid equations by magnetic resonance imaging in autosomal dominant
cystic kidney disease (ADPKD): executive summary from a Kidney Disease: polycystic kidney disease. Nephron 2015;129(4):253–262.
Improving Global Outcomes (KDIGO) Controversies Conference. Kidney 20. Lin YC, Lin CH, Lu HY, et al. Deep learning for fully automated tumor
Int 2015;88(1):17–27. segmentation and extraction of magnetic resonance radiomics features in
2. Kistler AD, Poster D, Krauer F, et al. Increases in kidney volume in autosomal cervical cancer. Eur Radiol 2020;30(3):1297–1305.
dominant polycystic kidney disease can be detected within 6 months. Kidney 21. Ge C, Gu IYH, Jakola AS, Yang J. Deep semi-supervised learning for brain
Int 2009;75(2):235–241. tumor classification. BMC Med Imaging 2020;20(1):87.
3. Chapman AB, Guay-Woodford LM, Grantham JJ, et al. Renal structure 22. Conze PH, Kavur AE, Cornec-Le Gall E, et al. Abdominal multi-organ
in early autosomal-dominant polycystic kidney disease (ADPKD): The segmentation with cascaded convolutional and adversarial deep networks.
Consortium for Radiologic Imaging Studies of Polycystic Kidney Disease Artif Intell Med 2021;117:102109.
(CRISP) cohort. Kidney Int 2003;64(3):1035–1045. 23. Mongan J, Moy L, Kahn CE Jr. Checklist for Artificial Intelligence in Medical
4. Sedman A, Bell P, Manco-Johnson M, et al. Autosomal dominant Imaging (CLAIM): A Guide for Authors and Reviewers. Radiol Artif Intell
polycystic kidney disease in childhood: a longitudinal study. Kidney Int 2020;2(2):e200029.
1987;31(4):1000–1005. 24. Kline TL, Korfiatis P, Edwards ME, et al. Performance of an Artificial
5. Fick-Brosnahan GM, Belz MM, McFann KK, Johnson AM, Schrier RW. Multi-observer Deep Neural Network for Fully Automated Segmentation
Relationship between renal volume growth and renal function in autosomal of Polycystic Kidneys. J Digit Imaging 2017;30(4):442–448.
dominant polycystic kidney disease: a longitudinal study. Am J Kidney Dis 25. Sharma K, Rupprecht C, Caroli A, et al. Automatic Segmentation of Kidneys
2002;39(6):1127–1134. using Deep Learning for Total Kidney Volume Quantification in Autosomal
6. Yu ASL, Shen C, Landsittel DP, et al. Baseline total kidney volume and Dominant Polycystic Kidney Disease. Sci Rep 2017;7(1):2049.
the rate of kidney growth are associated with chronic kidney disease pro- 26. Yakubovskiy P. Segmentation Models Pytorch. https://github.com/qubvel/
gression in Autosomal Dominant Polycystic Kidney Disease. Kidney Int segmentation_models.pytorch. Published 2020. Accessed May 2021.
2018;93(3):691–699. 27. Liu L, Jiang H, He P, et al. On the variance of the adaptive learning rate
7. Irazabal MV, Rangel LJ, Bergstralh EJ, et al. Imaging classification of autosomal and beyond. arXiv:1908.03265 [preprint] https://arxiv.org/abs/1908.03265.
dominant polycystic kidney disease: a simple model for selecting patients for Posted August 8, 2019. Accessed May 2021.
clinical trials. J Am Soc Nephrol 2015;26(1):160–172.

Radiology: Artificial Intelligence Volume 4: Number 2—2022  n  radiology-ai.rsna.org 7