Source-linked AI summary
The KiTS19 Challenge Data: 300 Kidney Tumor Cases with Clinical Context, CT Semantic Segmentations, and Surgical Outcomes
Nicholas Heller, Niranjan Sathianathen, Arveen Kalapara, Edward Walczak, Keenan Moore, Heather Kaluzniak, Joel Rosenberg, Paul Blake, Zachary Rengel, Makinna Oestreich, Joshua Dean, Michael Tradewell, Aneri Shah, Resha Tejpaul, Zachary Edgerton, Matthew Peterson, Shaneabbas Raza, Subodh Regmi, Nikolaos Papanikolopoulos, Christopher Weight
TL;DR
Quantitative kidney-tumor morphometry and treatment comparisons require more reliable imaging-derived details. The paper releases imaging, segmentation labels, and clinical information to support nephrometry research and automated segmentation, while using a kidney-labeling convention that avoids ambiguous hilar boundaries.
Problem
Existing nephrometry scores use relatively simple image features despite associations with surgical approach and multiple clinical outcomes, motivating richer tumor-detail quantification.
Method
The dataset combines manually reviewed clinical information, preoperative CT collection, and kidney and tumor segmentation annotations using thresholding and a defined hilar-labeling convention.
Results
The released data include imaging and segmentation labels for 210 public cases, with clinical attributes stored alongside each case.
Takeaways & Limitations
The release is intended to accelerate new nephrometric features and enable reliable learning-based semantic segmentation that can automate those features at scale.
Takeaways & Limitations
Intra-hilar boundaries were not reliably delineable, so intra-hilar structures were included in the kidney label to avoid ambiguity and annotation error.
Abstract
from arXiv · showhide
The morphometry of a kidney tumor revealed by contrast-enhanced Computed Tomography (CT) imaging is an important factor in clinical decision making surrounding the lesion's diagnosis and treatment. Quantitative study of the relationship between kidney tumor morphology and clinical outcomes is difficult due to data scarcity and the laborious nature of manually quantifying imaging predictors. Automatic semantic segmentation of kidneys and kidney tumors is a promising tool towards automatically quantifying a wide array of morphometric features, but no sizeable annotated dataset is currently available to train models for this task. We present the KiTS19 challenge dataset: A collection of multi-phase CT imaging, segmentation masks, and comprehensive clinical outcomes for 300 patients who underwent nephrectomy for kidney tumors at our center between 2010 and 2018. 210 (70%) of these patients were selected at random as the training set for the 2019 MICCAI KiTS Kidney Tumor Segmentation Challenge and have been released publicly. With the presence of clinical context and surgical outcomes, this data can serve not only for benchmarking semantic segmentation models, but also for developing and studying biomarkers which make use of the imaging and semantic segmentation masks.
1 Background & Summary
Kidney tumor nephrometry is clinically relevant but difficult to quantify at scale because existing scores use limited imaging features and annotated data are scarce. The KiTS19 dataset addresses this gap by combining imaging, segmentation labels, and clinical outcomes.
- Kidney tumor morphology from contrast-enhanced CT informs diagnosis and treatment decisions, but quantitative outcome studies are constrained by data scarcity and laborious feature measurement.
- Existing nephrometry systems, including R.E.N.A.L., P.A.D.U.A., and the Centrality Index, use cross-sectional imaging features and are associated with surgical approach and clinical outcomes.
- The dataset was released to accelerate development of new nephrometric features and reliable learning-based kidney and tumor semantic segmentation methods.
2 Methods
The dataset was assembled through retrospective clinical review, CT selection, manual annotation, and quality assurance for patients treated with nephrectomy. Inclusion required suitable preoperative late-arterial CT imaging and excluded cases that prevented unambiguous tumor-voxel definition.
- A retrospective review identified 300 eligible patients from 544 nephrectomy cases after CT availability and tumor-thrombus exclusions.
- The collection workflow comprised chart review, CT collection, CT annotation, and quality assurance, performed primarily by medical students supervised by a fellowship-trained urologic oncologist.
- Chart Review: Clinical attributes covering demographics, comorbidities, intervention, and outcomes were collected through electronic medical-record review and structured database queries.
- CT Collection: Each patient required a recent preoperative CT study containing a late-arterial contrast series depicting the entire abdomen, excluding cases with unavailable imaging or MRI-based planning.
- Annotations were distributed through a web application, drawn in the transverse plane, and longitudinally subsampled to roughly 50 kidney-containing slices per patient.
Manual Delineation
Manual delineation represents kidneys and kidney-tumor boundaries with contours, thresholding, and explicit handling of hilar structures. This design reduces voxel-wise annotation burden while producing a complete representation of the target boundaries.
- Annotators first confirmed complete kidney coverage, including horseshoe and transplanted kidneys, and represented each case as transverse slices.
- Kidney contours included the renal capsule, tumors, and cysts while excluding nonrenal radiodense tissue and intra-hilar ureteral and vascular structures.
- Each contour’s enclosed voxels were represented as A(i)j, while tumor contours were separately defined as C(i)j and excluded kidney tissue.
- Tumor contours specified the kidney-tumor interface because the remaining tumor boundary was already defined by the kidney contour.
- The contour-based procedure produced a complete and unambiguous kidney-tumor boundary representation while limiting tedious voxel-wise decisions.
- A 3x3 mean filter and a -30.0 HU threshold were used to separate perinephric fat from renal tissue, with a 7x7 median filter and 0 HU threshold used in selected noisy cases without cysts.
Thresholding and Hilum Filling
The procedure partitions thresholded, manually delineated slices into loose background, true tumor, and strict kidney bins, then identifies excluded kidney or cyst voxels to form final labels. Because intra-hilar boundaries are unreliable, these structures are included in the kidney label using a computed concavity boundary.
- Thresholding: Three voxel bins are formed between manual contours and thresholding: loose background, true tumor, and strict kidney.Loose background is outside the intersection of thresholded voxels and kidney-plus-tumor contour interiors; true tumor intersects tumor contours with kidney contours; strict kidney excludes tumor interiors.
- Thresholding: Strict kidney consists of thresholded voxels inside kidney contours but outside tumor contour interiors.
- Thresholding: Excluded kidney or cyst voxels are identified so the final background, kidney, and tumor partition can be computed for each annotated slice.
- Hilum Filling: Unreliable delineation of complex intra-hilar structures led the authors to include those structures in the kidney label rather than introduce ambiguous boundaries.The authors state that attempting to delineate this boundary would introduce ambiguity and error into the dataset.
- Hilum Filling: The hilum boundary is computed from the concavity in the manual contour using convexHull() and convexityDefects(), with heuristic selection followed by manual checking and correction.
- Hilum Filling: The algorithm produces final kidney and tumor labels by filling the hilum with thresholded tissue above the selected threshold.
Interpolation
For slices without manual contours, the method interpolates contours from neighboring annotated slices using contour-point matching and weighted coordinate interpolation. Contours are then processed as manually provided contours to compute ground-truth labels.
- Interpolation: Contours for unannotated slices are inferred from contours on slices l+a and l−b using Algorithm 1.
- Interpolation: After inferred contours are produced, ground truth is computed using the same procedure as for manually provided contours.
- Contour Matching: Algorithm 1 matches each contour point to the nearest point in the other contour, using centroid distance to guide matching.
- Contour Matching: Contour pairs are morphed only when their centroid distance is within the maximum distance Dmax = 20.
2.4 Code Availability
The authors note that code for hilum filling and interpolation should be made available through a GitHub repository with a runnable demonstration.
- Code Availability: The planned code release covers hilum filling and interpolation and includes a runnable demo.
2.5 Quality Assurance
Chart review used staged uncertainty resolution, while annotators were instructed to consult preoperative CT radiology notes to support their annotations.
- Chart Review: Uncertain chart-review fields were revisited by two students, with Christopher Weight resolving disagreements.
- Annotation: Students were instructed to read the preoperative CT radiology note before performing annotations.
Imaging Annotations
Tumor annotations were reviewed against expert clinician interpretation, with a second review of every image–ground-truth pair in transverse and coronal planes.
- A reviewing student checked every image–ground-truth pair in transverse and coronal planes for boundary consistency and concordance with the radiologist’s impression.
- Minor annotation issues were corrected directly by the reviewing student before acceptance.
- Rare major issues were returned to the first student for correction and subsequent review.
3 Data Records
The dataset stores each case in a numbered directory containing imaging, segmentation labels, and clinical attributes spanning demographics, treatment, outcomes, renal function, and pathology.
- The first 210 cases, numbered case 00000 through case 00209, comprise the public dataset portion.Cases use five-digit identifiers following the “case ” prefix.
- Each case directory contains imaging.nii.gz, segmentation.nii.gz, and a JSON file of clinical attributes.
- Clinical records include demographics, comorbidities, smoking and alcohol history, and preoperative and postoperative eGFR measurements.
- Operative and postoperative fields include surgery type, surgical procedure, approach, operative time, blood loss, complications, hospitalization, readmission, and emergency visits.
- Pathology fields record tumor size, malignancy, TNM stages, histologic subtype, necrosis, ISUP grade, and positive resection margins.
4 Technical Validation
Technical validation assessed the fidelity of semantic segmentation labels by repeating annotation on 30 randomly selected training cases and measuring agreement with the original labels.
- Agreement was measured using the average Sørenson Dice score.
- Table 1 reports this agreement over the 30 sampled cases.
5 Usage Notes
The release includes Python starter code for loading and visualizing the KiTS19 data.
- Python starter scripts were released to load and visualize the dataset.