PaperlayerView exact revision on arXiv

Source-linked AI summary

From Adult Renal Masses to Pediatric Wilms Tumors: Evaluating Cross-Domain Generalization of Kidney Lesion Segmentation on MRI

Sarah de Boer, Myrthe A. D. Buser, Hartmut Häntze, Jawed Nawabi, Lisa C. Adams, Alida F. W. van der Steeg, Marry M. van den Heuvel-Eibrink, Annemieke S. Littooij, Keno K. Bressem, Mathias Prokop, Alessa Hering

arXiv:2610.23568v1cs.CV

TL;DR

Pediatric renal tumor segmentation is limited by scarce annotated datasets, leaving open whether larger adult-trained models can generalize to pediatric Wilms tumors. The study evaluates an adult MRI segmentation model on adult and pediatric cohorts and compares it with a pediatric-specific model. The adult model generalized to the pediatric cohort, but its median tumor DSC was lower and non-inferiority was not demonstrated.

  • Problem

    Annotated pediatric oncology imaging datasets are scarce, and whether adult-trained MRI models generalize to pediatric Wilms tumors remains an open question.

    Source: §1 · p. 2

  • Method

    The study trained an adult kidney MRI segmentation model, evaluated it on held-out adult and pediatric Wilms tumor cohorts, and compared pediatric performance with a pediatric-specific nnU-Net.

    Source: §1 · p. 2, §2.3 · p. 4

  • Results

    Median tumor DSC was 0.90 (IQR: 0.86 – 0.92) for the pediatric-specific model versus 0.85 (IQR: 0.64 – 0.89) for the adult model, and non-inferiority was not demonstrated.

    Source: §3.3 · p. 5

  • Takeaways & Limitations

    The adult model demonstrated transferability to pediatric Wilms tumor MRI, but the performance gap indicates adult datasets cannot fully compensate for pediatric-specific differences.

    Source: §4 · p. 7

    The model comparison was asymmetric, the pediatric cohort was relatively small, and acquisition and annotation differences prevented attributing the performance gap to a single source of domain shift.

    Source: §4 · p. 6

Abstract

from arXiv · show

Deep learning segmentation models for pediatric oncology are often trained on small, single-center cohorts, whereas much larger datasets exist for adult diseases, such as kidney cancer. Whether models trained on adult renal lesions generalize to pediatric Wilms tumors remains unclear. We trained a segmentation model on single-sequence T1- or T2-weighted adult kidney MRI and evaluated it on a held-out adult test set and a pediatric Wilms tumor cohort, where it was compared with a pediatric-specific nnU-Net. Segmentation performance was quantified using the Dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance. On the adult test set, the model achieved a median DSC of 0.86 for renal lesions. On the pediatric cohort, the adult-trained model achieved a median DSC of 0.85 compared with 0.90 for the pediatric-specific model. Despite differences in age, anatomy, and tumor type, the adult-trained model showed promising performance on pediatric Wilms tumors, with only a modest difference in median DSC compared with the pediatric-specific model. However, non-inferiority could not be shown, highlighting the value of pediatric training data. In future work, a fine-tuning strategy could be explored.

1 Introduction

Pediatric renal tumor segmentation is constrained by scarce annotated imaging data, while adult renal disease cohorts are substantially larger. This study tests whether an adult-trained MRI model generalizes to pediatric Wilms tumors and compares it with a pediatric-specific model.

  • Pediatric oncology segmentation supports tumor volume assessment, treatment planning, and response monitoring, but annotated imaging datasets remain scarce compared with adult renal disease cohorts.Source: §1 · p. 2
  • Whether adult-trained models generalize to pediatric MRI lesion segmentation is an open question not previously investigated, and no publicly available adult kidney cancer MRI model had enabled direct assessment.Source: §1 · p. 2
  • The study evaluates an adult MRI model on adult hold-out and pediatric Wilms tumor data, compares it per case with a pediatric-specific nnU-Net, and releases the adult model for further evaluation.Source: §1 · p. 2

2 Materials and Methods

The study uses adult MRI data for training and held-out testing, then evaluates transfer to a pediatric Wilms tumor cohort. A semi-supervised workflow addresses missing contralateral kidney annotations, and evaluation includes comparison with a pediatric-specific model.

2.1 Data

  • The adult cohort comprised 1,794 MRI sequences from 689 patients, with 20% held out; the pediatric test cohort comprised 45 patients with Wilms tumor and manually reviewed lesion annotations.Source: §2.1 · pp. 2–3

2.2 Model training

  • Because adult reference annotations omitted contralateral structures, MRSegmentator localized kidneys, a temporary nnU-Net generated pseudolabels, and merged manual and predicted masks trained the final model.Source: §2.2 · p. 3, Figure 1 · p. 3

2.3 Experimental design and evaluation

  • The evaluation measured DSC and HD95, tested adult-model performance on held-out adult and pediatric data, and assessed pediatric non-inferiority using a paired Wilcoxon test with a 0.03 margin.Adult-to-pediatric differences were reported descriptively because cohort differences could independently affect segmentation metrics.Source: §2.3 · pp. 4–5

3 Results

The adult model performed well on held-out adult renal masses and generalized to pediatric Wilms tumors, but the pediatric-specific model had higher median tumor DSC and adult-model non-inferiority was not demonstrated.

3.1 Adult in-domain performance

  • On the adult hold-out set, renal mass segmentation achieved median DSC 0.86 (IQR: 0.79 – 0.91) and median HD95 6.3 mm (IQR: 4.0 – 15.6).Median renal mass DSC was 0.87 for T1-weighted fat-suppressed and T2-weighted sequences, and 0.84 for conventional T1-weighted sequences.Source: §3.1 · p. 5

3.2 Generalization gap

  • On the pediatric cohort, the adult model achieved a median tumor DSC of 0.85 (IQR: 0.64 - 0.89) and median HD95 of 11.0 mm (IQR: 6.8 - 16.8).Five patients had DSC 0, including three cases with no kidney or tumor segmentation; all three were younger than 10 months.Source: §3.2 · p. 5

3.3 Comparison with pediatric model

  • Median tumor DSC was 0.90 (IQR: 0.86 – 0.92) for the pediatric-specific model versus 0.85 (IQR: 0.64 – 0.89) for the adult model; non-inferiority was not demonstrated.Median HD95 was 6.6 mm (IQR: 4.9 – 11.8) and 11.0 mm (IQR: 6.8 – 16.8), respectively.Source: §3.3 · p. 5

4 Discussion and conclusion

Although the adult model generalized to Wilms tumor MRI, a performance gap remained relative to pediatric-specific modeling, and study design limits direct interpretation of that difference.

4 Discussion and conclusion

  • Renal lesion DSC was 0.86 and kidney DSC was 0.92; comparison with MRSegmentator’s bilateral-kidney DSC of up to 0.95 is limited by evaluation of tumor-affected kidneys only.Source: §4 · p. 5
  • The adult model was not trained on Wilms tumors or pediatric data; tumor segmentation was absent in four cases, and performance did not systematically increase with patient age.Source: §4 · pp. 5–6
  • Wilms tumor morphology and smaller pediatric anatomy may contribute to the performance gap, while fixed isotropic resampling can reduce effective resolution of small structures.Source: §4 · p. 6
  • The adult-versus-pediatric comparison was asymmetric, and differences in cohort size, acquisition, and annotation make the performance gap difficult to attribute to one source.Source: §4 · p. 6
  • Future work could test domain adaptation or targeted fine-tuning to improve transfer while reducing reliance on large pediatric training cohorts.Source: §4 · p. 7
  • The adult model generalized to pediatric Wilms tumor MRI, but the performance gap indicates that large adult datasets cannot fully compensate for pediatric-specific differences.The authors suggest adult datasets may still support pediatric annotation and model development.Source: §4 · p. 7
Loading 2610.23568v1…