EFNet: evidence fusion network for tumor segmentation from PET-CT volumes

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Precise tumor delineation from PET-CT is crucial for radiation therapy.
• Current PET-CT co-segmentation methods using fully convolutional neural networks (FCNs) often employ complex fusion strategies that do not account for multi-modal uncertainty and are computationally intensive.
• Existing methods struggle with uncertainty in single-modal segmentation and high computational costs for 3D volumes.

Purpose of the Study:

• To develop a novel PET-CT co-segmentation method that addresses the limitations of current fusion strategies by incorporating uncertainty.
• To propose the Evidence Fusion Network (EFNet) that leverages uncertainty quantification for improved segmentation accuracy and computational efficiency.
• To enhance tumor delineation for more effective clinical practice and radiation therapy.

Main Methods:

• Proposed the Evidence Fusion Network (EFNet), utilizing a 3D U-Net backbone with unidirectional feature fusion.
• Introduced an 'evidence loss' function to generate PET and CT segmentation results with associated uncertainty (evidence).
• Implemented an evidence fusion mechanism to combine PET and CT evidence, reducing uncertainty and improving segmentation.
• EFNet allows for separate training and prediction of PET and CT evidence, simplifying the network architecture.

Main Results:

• EFNet demonstrated significant improvements in segmentation accuracy on soft-tissue-sarcomas and lymphoma datasets.
• Achieved an 8% and 5% increase in Dice score compared to 3D U-Net for the respective datasets.
• Outperformed complex feature fusion methods by 7% and 2% in Dice score.
• The proposed method simplifies the network while enhancing segmentation performance.

Conclusions:

• Outputting uncertainty evidence and employing evidence fusion in FCN-based PET-CT segmentation simplifies the network and improves results.
• EFNet offers a more efficient and accurate approach to PET-CT co-segmentation compared to existing methods.
• The findings suggest a promising direction for enhancing tumor delineation in clinical applications and radiation therapy planning.