Stress-testing cross-cancer generalizability of 3D nnU-Net for PET-CT tumor segmentation: multi-cohort evaluation with novel oesophageal and lung cancer datasets

2025-08-25

AI Summary

This study addresses the limited generalizability of deep learning models for clinical PET-CT tumor segmentation across cancer types, imaging devices, and patient populations. To systematically evaluate robustness, we adopt 3D nnU-Net and propose the first multi-center stress-testing framework grounded in population and oncologic diversity. It integrates newly curated esophageal and lung cancer datasets with the public AutoPET dataset. We comparatively assess three training paradigms: single-cancer training, public-dataset-only training, and multi-cancer joint training. Results demonstrate that joint training substantially improves cross-domain generalization—achieving Dice Similarity Coefficients of 40.7% (esophageal), 52.9% (lung), and 60.9% (AutoPET)—with the most pronounced reduction in boundary localization error. Crucially, data diversity, rather than model complexity, emerges as the primary driver of clinical robustness. Our findings establish multi-cancer joint training as a core strategy for enhancing domain-agnostic performance in real-world PET-CT segmentation.

Technology Category

Machine Learning: Transfer, Domain Adaptation, Multi-Task LearningComputer Vision: SegmentationIntelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

Responsible Web: Machine-in-the-loop, human agency and autonomySearch and Retrieval-Augmented AI: Vertical and domain-specific searchGraph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Robust generalization is essential for deploying deep learning based tumor segmentation in clinical PET-CT workflows, where anatomical sites, scanners, and patient populations vary widely. This study presents the first cross cancer evaluation of nnU-Net on PET-CT, introducing two novel, expert-annotated whole-body datasets. 279 patients with oesophageal cancer (Australian cohort) and 54 with lung cancer (Indian cohort). These cohorts complement the public AutoPET dataset and enable systematic stress-testing of cross domain performance. We trained and tested 3D nnUNet models under three paradigms. Target only (oesophageal), public only (AutoPET), and combined training. For the tested sets, the oesophageal only model achieved the best in-domain accuracy (mean DSC, 57.8) but failed on external Indian lung cohort (mean DSC less than 3.4), indicating severe overfitting. The public only model generalized more broadly (mean DSC, 63.5 on AutoPET, 51.6 on Indian lung cohort) but underperformed in oesophageal Australian cohort (mean DSC, 26.7). The combined approach provided the most balanced results (mean DSC, lung (52.9), oesophageal (40.7), AutoPET (60.9)), reducing boundary errors and improving robustness across all cohorts. These findings demonstrate that dataset diversity, particularly multi demographic, multi center and multi cancer integration, outweighs architectural novelty as the key driver of robust generalization. This work presents the demography based cross cancer deep learning segmentation evaluation and highlights dataset diversity, rather than model complexity, as the foundation for clinically robust segmentation.

Problem

• Evaluating nnU-Net's cross-cancer generalization for PET-CT tumor segmentation
• Testing performance across diverse anatomical sites and patient demographics
• Assessing impact of dataset diversity on segmentation robustness

Innovation

• Used 3D nnU-Net for PET-CT tumor segmentation
• Trained models with multi-cohort multi-cancer datasets
• Combined approach achieved balanced cross-domain performance