MO012: Development of an Accurate Automated Segmentation Algorithm to Measure Total Kidney Volume in ADPKD Suitable for Clinical Application (The Cystvas Study). (3rd May 2022)

• Record Type: Journal Article
• Authors: Taylor, Jonathan; Thomas, Richard; Metherall, Peter; Ong, Albert; Simms, Roslyn
• Is Part Of: Nephrology dialysis transplantation. Volume 37(2022)Supplement 3
• Publication Date: 2022-05-03
• Subjects: Nephrology -- Periodicals; Hemodialysis -- Periodicals; Kidneys -- Transplantation -- Periodicals; Hemodialysis; Kidneys -- Transplantation; Nephrology; Periodicals; 616.61
• DOI: 10.1093/ndt/gfac061.007
• Languages: English
• ISSNs: 0931-0509

Abstract

Abstract: BACKGROUND AND AIMS: A major barrier to the routine adoption of total kidney volume (TKV) as a clinical biomarker of disease for autosomal dominant polycystic disease (ADPKD) is the significant human operator time required even by experienced analysts (typically, 45–90 min per patient). Several groups have investigated automated and semi-automated kidney segmentation methods to either reduce or eliminate the human interaction required. However, such tools have mostly been developed using data from single centers, which may not translate well to other centers. To date, there has been little attempt to develop or validate algorithms using multi-center and multi-scanner data. Here, we report an automated segmentation tool capable of high performance across different patient populations and scanner sequences using 1.5 T MRI data from four centers (the CYSTic consortium). The algorithm was subsequently tested in a separate clinical cohort to assess its likely performance during routine clinical use. METHOD: All 1.5 T studies from the CYSTic trial were downloaded (acquired from Siemens Avanto, GE Optima and Siemens Aera, using different sequences). Cases with poor image quality or with sections of kidney missing from the field of view were excluded. A single, experienced operator selected the most appropriate image series for segmentation and manually segmented each patient's kidneys using a commercial software program (MIM Encore). There were 454 kidneys segmented from 227 scans. These data were used for algorithm training and validation. In addition, 48 routine clinical scans from the Sheffield 3D Lab were extracted from the archives along with their original segmentations (performed by six different analysts), to use as a test set. None of the patients in the clinical test set were included in the training set. An ensemble U-net algorithm was created using the nnUNet approach Isensee et al. (nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Meth 2021; 18 (2): 203–211), whereby the CYSTic data were used in a 5-fold cross-validation, with stratification across the four centers (i.e. each center contributed 80% of the available data to algorithm training in each fold). Algorithm training proceeded according to the standard heuristic nnUnet functions, using a 3D architecture, for 100 epochs. Segmented kidneys were split into left and right sides during post-processing, through analysis of the position of the center of gravity of segmented regions. Once trained, the five algorithms from cross-validation were applied in an ensemble to the clinical test cohort. RESULTS: In both cross-validation and clinical testing phases, the median DICE score was 0.96 for each kidney (IQR of 0.95–0.97 in cross-validation on both sides, And 0.95–0.97 on the left side for clinical testing and 0.96 for the right). The median total kidney volume error was −0.46% (−2.02 to 1.27) for the left side in cross-validation and −0.82% (−2.55 to 0.86) for the right. In the clinical testing phase, the median volume errors were −1.8% (−3.69 to 1.29), left and −1.79% (−3.95 to 0.65), right. The mean time taken to manually segment kidneys in the CYSTIc dataset was 54 min per scan (SD of 31 min). Use of the algorithm as a first pass segmentation, with subsequent checking and editing by an operator, would significantly reduce human input time to a few minutes per case CONCLUSION: Our new algorithm demonstrates high accuracy compared to the gold standard of manual TKV segmentation and performs well in a wide range of patients with ADPKD imaged using different scanners at several European centers. Its high performance in a real-world clinical dataset demonstrates that such tools can provide a reliable means of measuring TKV in routine practice and reduces the previous barrier of analyst time and experience.