Research map / Detection and segmentation / Open-set and few-shot detection

PanTS: The Pancreatic Tumor Segmentation Dataset

NeurIPS 2025 · Detection and segmentation › Open-set and few-shot detection · approach: Remote sensing segmentation and change · 7 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces PanTS, a 36,390-CT multi-institutional pancreatic tumor and 24-structure anatomy segmentation dataset, and shows that nnU-Net trained on it beats MSD/PANORAMA-trained models on out-of-distribution tumor detection and segmentation.

Problem

Public pancreatic CT datasets are small, homogeneous, and often lack tumor annotations for non-PDAC lesions and anatomical context, so AI models generalize poorly to diverse clinical settings.

Approach

A very large multi-center pancreatic CT dataset with voxel-wise tumor annotations plus pancreas head/body/tail and 24 surrounding structures, with an out-of-distribution test set from unseen centers; the paper argues that tumor annotation scale and anatomical context each improve tumor segmentation.

Key results

• nnU-Net trained on PanTS reaches AUC 0.959 for tumor detection on the held-out PanTS test set, versus 0.819 with PANORAMA and 0.810 with MSD-Pancreas training.
• The 28-class model (tumor, pancreas subregions, 24 structures) improves mean tumor DSC from 57.4% to 67.7% over a 2-class model (+10.3%).
• PanTS-trained nnU-Net outperforms prior MSD-trained methods by at least +4.9% DSC and +3.1% NSD on the official MSD-Pancreas test set and ranks top-1 on that leaderboard.

Measured effects reported in the paper

Surprising findings

• Going from MSD-Pancreas (n=281) to PANORAMA (n=2,238) barely improved AUC (0.810 to 0.819), showing scale alone is not sufficient without comprehensive tumor annotation.
• Inter-annotator agreement on tumor/pancreas labels had a median DSC of only 86.1% with substantial low-agreement cases, and the authors note even experts can miss early tumors.

Limitations

• Test set is not publicly released and is reserved for third-party evaluation, limiting open reproduction of test results.
• Training set is partly assembled from existing public datasets with only 1,076 tumor annotations, and tumor labels are described as silver-standard.
• Test set differs substantially from training set in age, sex, slice thickness, and contrast phase, so the OOD gains may be confounded by distribution shift.

Experimental setup

Datasets: PanTS training set MSD-Pancreas PANORAMA BTCV AbdomenCT-1K AMOS WORD CHAOS KiTS'23 LiTS TCIA Pancreas-CT CT-ORG

Benchmarks: PanTS test set MSD Pancreas leaderboard

Metrics: DSC NSD AUC Sensitivity Specificity

Baselines: Kim et al. NAS nnU-Net (MSD) C2FNAS DiNTS Swin UNETR Universal Model PANORAMA-trained nnU-Net MSD-Pancreas-trained nnU-Net

Builds on: nnU-Net MONAI-Label

Closest papers by meaning

• ZePT: Zero-Shot Pan-Tumor Segmentation via Query-Disentangling and Self-Prompting (CVPR 2024 · similarity 0.82)
• Cross-Modal Interactive Perception Network with Mamba for Lung Tumor Segmentation in PET-CT Images (CVPR 2025 · similarity 0.82)
• Touchstone Benchmark: Are We on the Right Way for Evaluating AI Algorithms for Medical Segmentation? (NeurIPS 2024 · similarity 0.82)
• STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology (NeurIPS 2025 · similarity 0.82)
• Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · similarity 0.82)
• "Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence (ICML 2025 · similarity 0.81)