Open-Source PET/CT Foundation Model for Tumor Segmentation

A new open-source foundation model for whole-body FDG PET/CT tumor segmentation has been developed using 4,997 harmonized scans from four public datasets. The model employs hierarchical UNet-shaped backbones with early channel-wise concatenation to enable cross-modal interaction between CT and PET from the first embedding layer. It introduces a masked autoencoding objective based on zero-mean imputation combined with a weighted global reconstruction loss. The approach aims to overcome limitations of existing deep learning methods that are task-specific, single-center, or use dual-branch fusion schemes that delay cross-modal interaction. The model is designed for oncologic imaging, leveraging synergistic interpretation of anatomical (CT) and metabolic (PET) information.