Anatomy consistent segmentation network for joint PET/CT tumor segmentation

Joint PET/CT imaging combines functional and anatomical information, making multimodal tumor segmentation essential in clinical oncology. However, segmentation performance is often hindered by inter-modal feature entanglement and the lack of reliable uncertainty estimates. To tackle these issues, we propose an Anatomy-Consistent Segmentation Network (ACSN) based on feature disentanglement. In this network, the multimodal inputs are first processed with a dual-branch adversarial Feature Disentanglement Module (FDM) to extract the shared anatomical features of PET and CT. These features are then fed into a Feature Fusion Module (FFM) to extract tumor-specific representations via two Multi-receptive field Segmentation Backbones (MSB) (for CT and PET, respectively) and to extract fused features (from PET and CT) based on their uncertainties. In each MSB, a Feature Space Adjustment (FSA) module is developed to better align anatomical and tumor-related features. To better fuse multi-modal features and to mitigate the mis-calibration risk of traditional uncertainty calibration approaches, a Layer-wise Uncertainty Calibrator (LUC) is developed to progressively calibrate multi-scale features across the network, enabling finer-grained modeling of epistemic uncertainty and guiding more reliable multimodal fusion. The final tumor prediction is trustworthily produced by integrating calibrated evidence from PET, CT, and their fused representation under the guidance of Dempster-Shafer theory (DST). Experiments on the AutoPET and Hecktor datasets show that ACSN achieves competitive segmentation performance under the adopted 2D axial slice-wise evaluation protocol and provides calibrated uncertainty estimates that are associated with segmentation errors. These results suggest that ACSN may provide auxiliary information for uncertainty-aware PET/CT tumor segmentation and quality control. Our code will be available at https://github.com/narutodd/ACSN.