Anatomy consistent segmentation network for joint PET/CT tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Multimodal PET/CT imaging is crucial for oncology, but tumor segmentation faces challenges with feature entanglement and unreliable uncertainty estimates.
• Existing methods struggle to effectively integrate anatomical (CT) and functional (PET) data while quantifying segmentation confidence.

Purpose of the Study:

• To develop an Anatomy-Consistent Segmentation Network (ACSN) that enhances multimodal tumor segmentation by addressing feature entanglement and improving uncertainty estimation.
• To provide reliable uncertainty quantification for PET/CT tumor segmentation to aid clinical decision-making.

Main Methods:

• Proposes an ACSN utilizing a dual-branch adversarial Feature Disentanglement Module (FDM) to separate shared anatomical features from PET and CT data.
• Incorporates a Feature Fusion Module (FFM) with Multi-receptive field Segmentation Backbones (MSB) and a Layer-wise Uncertainty Calibrator (LUC) for calibrated multimodal fusion.
• Employs Dempster-Shafer theory (DST) for integrating calibrated evidence from PET, CT, and fused representations to produce final tumor predictions.

Main Results:

• ACSN demonstrates competitive 2D axial slice-wise segmentation performance on the AutoPET and Hecktor datasets.
• The network provides calibrated uncertainty estimates that correlate with segmentation errors, indicating improved reliability.
• The proposed Layer-wise Uncertainty Calibrator (LUC) effectively mitigates mis-calibration risks in uncertainty estimation.

Conclusions:

• ACSN offers a promising approach for uncertainty-aware PET/CT tumor segmentation in clinical oncology.
• The method's ability to disentangle features and calibrate uncertainty can enhance diagnostic accuracy and quality control.
• The developed ACSN framework has the potential to serve as valuable auxiliary information for oncologists using PET/CT imaging.