Robust Kidney Abnormality Segmentation: A Validation Study of an AI-Based Framework

TL;DR

The study tackles the lack of robust, generalizable kidney abnormality segmentation by training a nnU-Net-based framework on two public datasets and validating on three external test sets totaling over 1,500 CT scans. It employs a preprocessing step with TotalSegmentator and careful post-processing, achieving superior segmentation and detection performance compared with TotalSegmentator and BAMF across diverse datasets and subgroups. The results show Dice improvements and competitive $HD_{95}$ distances, with performance approaching inter-observer variability for healthy kidneys and abnormalities on external data, underscoring clinical relevance. Public availability of the model and code supports immediate translation and further cross-institutional evaluation for volume-based biomarkers and automated kidney diagnostics.

Abstract

Kidney abnormality segmentation has important potential to enhance the clinical workflow, especially in settings requiring quantitative assessments. Kidney volume could serve as an important biomarker for renal diseases, with changes in volume correlating directly with kidney function. Currently, clinical practice often relies on subjective visual assessment for evaluating kidney size and abnormalities, including tumors and cysts, which are typically staged based on diameter, volume, and anatomical location. To support a more objective and reproducible approach, this research aims to develop a robust, thoroughly validated kidney abnormality segmentation algorithm, made publicly available for clinical and research use. We employ publicly available training datasets and leverage the state-of-the-art medical image segmentation framework nnU-Net. Validation is conducted using both proprietary and public test datasets, with segmentation performance quantified by Dice coefficient and the 95th percentile Hausdorff distance. Furthermore, we analyze robustness across subgroups based on patient sex, age, CT contrast phases, and tumor histologic subtypes. Our findings demonstrate that our segmentation algorithm, trained exclusively on publicly available data, generalizes effectively to external test sets and outperforms existing state-of-the-art models across all tested datasets. Subgroup analyses reveal consistent high performance, indicating strong robustness and reliability. The developed algorithm and associated code are publicly accessible at https://github.com/DIAGNijmegen/oncology-kidney-abnormality-segmentation.

Figures (11)

• Figure 1: Overview of the research. The proposed AI-based segmentation model is trained on two publicly available datasets. The proposed framework includes a pre-processing step utilizing TotalSegmentator to crop out a region-of-interest around the kidney region. A trained nnU-Net segments the kidney and kidney abnormalities. We thoroughly validate the proposed method on three test datasets from three different medical centres, and test performance across patient sex, patient age, tumor histologic subtype and CT contrast phase.
• Figure 2: Visualization of the annotation protocols of the two training datasets. On the left, a segmentation mask from the KiTS dataset is shown in , which shows that the hilum is included in the annotation. On the right, a segmentation mask from the Radboudumc dataset is shown in , which shows that the hilum was left out of the kidney region while annotating. In the second row, we visualize that KiTS annotations distinguish between cysts and tumors and Radboudumc annotations only include an abnormality region .
• Figure 3: Segmentation results on the healthy cohort of the Radboudumc test set expressed in Dice (left) and Hausdorff distance (mm) at 95th percentile (right). We present the proposed model , Total Segmentator and the BAMF model . The boxplots show the median (horizontal line in the box), quartiles (presented by the box) and the full distribution (presented by the whiskers). For the purpose of readability we omitted the outliers from these plots.
• Figure 4: Segmentation results expressed in Dice (left column) and Hausdorff distance (mm) at 95th percentile (right column). The results are shown for three segmentation regions, Kidney, Kidney+Abnormality and Abnormality. We present results per dataset and show the proposed model , Total Segmentator and the BAMF-model . The boxplots show the median (horizontal line in the box), quartiles (presented by the box) and the full distribution (presented by the whiskers). For the purpose of readability we omitted the outliers from these plots.
• Figure 5: Subgroup analysis performed on the TCGA-KIRC test set. Subgroups that are tested for are patient sex and patient age. We show performance measured in Dice. The boxplots show the median (horizontal line in the box), quartiles (presented by the box) and the full distribution (presented by the whiskers). For the purpose of readability we omitted the outliers from these plots.