Joint Tumor Segmentation in PET-CT Images Using Co-Clustering and Fusion Based on Belief Functions

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate tumor segmentation is crucial for effective radiation therapy.
• Hybrid PET-CT (positron emission tomography-computed tomography) is standard, but many segmentation methods use only single modalities.
• Existing methods struggle with the inherent uncertainty and noise in PET-CT data.

Purpose of the Study:

• To develop a co-clustering algorithm for simultaneous 3D tumor segmentation in PET-CT images.
• To leverage the complementary functional (PET) and anatomical (CT) information for improved segmentation accuracy.
• To address uncertainty and noise in medical images using belief function theory.

Main Methods:

• A novel co-clustering algorithm is proposed for concurrent segmentation of PET and CT modalities.
• Belief function theory is employed to model, fuse, and reason with uncertain image data.
• Iterative adaptation of distance metrics and a context term encourage modality-specific reliability and cross-modality consistency.

Main Results:

• The co-clustering algorithm demonstrated robust performance in segmenting 3D tumors from PET-CT images.
• Evaluation on 21 non-small cell lung cancer patient datasets showed competitive results compared to state-of-the-art methods.
• The belief-functions-based fusion strategy enhanced segmentation consistency across modalities.

Conclusions:

• The proposed co-clustering method effectively integrates functional and anatomical information from PET-CT for precise tumor delineation.
• This approach offers a promising solution for improving radiation therapy planning by enhancing segmentation accuracy in the presence of image noise and uncertainty.
• The integration of belief functions provides a robust framework for handling imprecise medical imaging data.