Hyper-Connected Transformer Network for Multi-Modality PET-CT Segmentation.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• [18F]-Fluorodeoxyglucose (FDG) positron emission tomography - computed tomography (PET-CT) is crucial for cancer diagnosis.
• Automatic tumor segmentation and computer-aided diagnosis require co-learning complementary PET-CT imaging features.
• Existing methods may not fully leverage the complementary information present in multi-modality PET-CT images.

Purpose of the Study:

• To propose a novel hyper-connected transformer (HCT) network for enhanced feature learning from multi-modality PET-CT images.
• To improve automatic tumor segmentation and computer-aided cancer diagnosis systems.
• To integrate transformer networks (TN) with hyper-connected fusion for comprehensive image analysis.

Main Methods:

• Developed a hyper-connected transformer (HCT) network integrating transformer networks (TN) with hyper-connected fusion.
• Utilized TN with self-attention for global dependency and image-wide contextual information extraction.
• Extended TN with multiple branches for separate feature extraction and introduced iterative hyper-connected fusion for multi-modal data.

Main Results:

• The HCT network demonstrated superior performance in segmentation accuracy compared to existing methods.
• Evaluated on two clinical datasets, confirming the effectiveness of the proposed approach.
• The co-learning of complementary features significantly improved the analysis of PET-CT imaging.

Conclusions:

• The proposed HCT network offers a powerful approach for integrating multi-modality PET-CT imaging features.
• This method can serve as a valuable tool for physicians in tumor quantification and identifying imaging biomarkers.
• The HCT network advances the development of computer-aided diagnosis systems for improved cancer care.