Related Experiment Video
Updated: Jun 5, 2026
10:04
Analysis of 18FDG PET/CT Imaging as a Tool for Studying Mycobacterium tuberculosis Infection and Treatment in Non-human Primates
Published on: September 5, 2017
An Open Multi-Center Whole-Body FDG PET/CT Foundation Model for Tumor Segmentation
Xiaofeng Liu1, Qianru Zhang1, Thibault Marin1
1Department of Radiology and Biomedical Imaging, Yale Biomedical Imaging Institute, Yale University, New Haven, CT, USA.
Arxiv
|June 4, 2026
Summary
A new foundation model for PET/CT imaging integrates anatomical and metabolic data early for improved tumor segmentation. This open-source tool enhances deep learning efficiency, reducing the need for extensive manual annotations in oncology.
Area of Science:
- Oncologic Imaging
- Medical Image Analysis
- Deep Learning
Background:
- Synergistic interpretation of computed tomography (CT) and positron emission tomography (PET) is crucial for oncologic imaging.
- Current deep learning models for PET/CT are often task-specific, single-center, and delay cross-modal interaction.
Purpose of the Study:
- To develop an open-source, multi-center, whole-body FDG PET/CT foundation model.
- To improve early spatial correspondence and cross-modal interaction between PET and CT data.
- To enhance label efficiency and representation learning for PET/CT tumor segmentation.
Main Methods:
- Utilized 4,997 harmonized scans from four public datasets.
- Employed hierarchical UNet-shaped backbones with early channel-wise concatenation for feature interaction.
- Introduced a masked autoencoding objective with zero-mean imputation and weighted global reconstruction loss.
Main Results:
- Demonstrated strong label efficiency, achieving performance comparable to full-dataset training with only 10% labeled data on AutoPET lesion segmentation.
- Achieved higher Dice scores with joint PET/CT pretraining compared to separated-modality pretraining under 5-shot linear probing.
- Showcased effective cross-modality representation learning for PET/CT tumor segmentation.
Conclusions:
- The proposed multi-center foundation model offers significant label efficiency for PET/CT tumor segmentation.
- Provides a robust, open-source basis for advancing automated oncologic imaging.
- Reduces the clinical need for large-scale manual annotations in medical imaging.
