CoF-DResNet: Cancer Metastasis Recognition Network based on Dynamic Coordinated Metabolic Attention and Structural...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Cancer Diagnostics

Background:

• Cancer metastasis identification is crucial for patient prognosis and treatment planning.
• ¹⁸F -FDG PET/CT offers combined metabolic and structural imaging for cancer diagnosis.
• Integrating metabolic and structural information is key for improving metastasis recognition.

Purpose of the Study:

• To propose a novel cancer metastasis identification network.
• To enhance feature representation and recognition performance by integrating metabolic and structural information.
• To improve the efficacy of feature expression through multi-receptive field fusion.

Main Methods:

• A cancer metastasis identification network utilizing dynamic coordinated metabolic attention and structural attention.
• Incorporation of a dynamic coordinated attention module (DCAM) into ResNet branches to extract metabolic and structural features.
• Utilizing a multi-receptive field feature fusion module (MRFM) for semantic feature fusion.

Main Results:

• Experiments were conducted on private lung lymph nodes and public soft tissue sarcomas datasets.
• The proposed method achieved an accuracy of 76.0% on the lung lymph nodes dataset.
• The proposed method achieved an accuracy of 75.1% on the soft tissue sarcomas dataset.

Conclusions:

• The proposed network demonstrates improved accuracy compared to standard ResNet, with increases of 6.8% and 5.6% on the respective datasets.
• The integration of dynamic coordinated attention and multi-receptive field feature fusion enhances cancer metastasis identification.
• The study affirms the efficacy of the proposed method for advanced cancer diagnostic imaging.