Multi-Modal Co-Learning for Liver Lesion Segmentation on PET-CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Liver lesion segmentation is crucial for hepatocellular carcinoma (HCC) diagnosis and treatment planning.
• Multi-modal PET-CT scans offer complementary information but current segmentation methods fail to fully leverage cross-modal interactions and multi-resolution feature fusion.

Purpose of the Study:

• To develop an advanced model for liver lesion segmentation that effectively integrates information from multi-modal PET-CT scans.
• To address limitations in existing methods by enhancing cross-modal feature interaction and feature map fusion across different resolutions.

Main Methods:

• A novel model architecture is proposed featuring shared down-sampling blocks for cross-modal feature interaction between PET and CT encoding branches.
• The model combines feature maps from different resolutions to create spatially varying fusion maps, enhancing lesion information.
• A similarity loss function is introduced to ensure consistency between predictions from separate network branches.

Main Results:

• The proposed model was evaluated on a PET-CT dataset for liver tumor segmentation.
• Comparative analysis against baseline multi-modal segmentation techniques (multi-branches, multi-channels, cascaded networks) demonstrated superior performance.
• The method achieved significantly higher accuracy in liver lesion segmentation compared to existing approaches.

Conclusions:

• The developed model effectively addresses limitations in current multi-modal liver lesion segmentation by improving feature interaction and fusion.
• The proposed approach offers a significant advancement in accuracy for hepatocellular carcinoma diagnosis and treatment planning using PET-CT scans.