ICT-Net: An Integrated Convolution and Transformer-Based Network for Complex Liver and Liver Tumor Region...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of liver and hepatocellular carcinoma (HCC) from CT images is crucial for diagnosis and treatment.
• Deep learning, particularly transformers, shows promise for computer-aided diagnosis (CAD) but faces challenges in spatial feature extraction and limited annotated datasets.
• Existing models struggle with robust feature extraction for precise liver and lesion segmentation.

Purpose of the Study:

• To address limitations in liver tumor segmentation, this study introduces a novel deep learning architecture.
• To develop a more accurate method for segmenting liver regions and hepatocellular carcinoma (HCC) from computed tomography (CT) images.
• To overcome challenges in spatial feature extraction and limited annotated liver datasets for HCC.

Main Methods:

• A new liver dataset with HCC annotations (CCH-LHCC-CT) was created.
• A novel deep learning architecture, ICT-Net, was developed using a pretrained transformer encoder and an advanced decoder.
• ICT-Net incorporates feature upscaling and enhanced convolution-transformer blocks for improved segmentation.

Main Results:

• The ICT-Net model was evaluated on the CCH-LHCC-CT dataset and three public CT liver datasets.
• ICT-Net achieved superior segmentation accuracy, indicated by higher ACC, DSC, and IoU, and lower HD95 across all tested datasets.
• The proposed architecture demonstrated robust performance in segmenting both liver regions and liver tumors.

Conclusions:

• The developed ICT-Net model provides robust information for accurate liver and liver tumor segmentation.
• ICT-Net significantly outperforms existing methods in segmentation accuracy metrics (ACC, DSC, IoU).
• ICT-Net shows potential for clinical translation by enhancing surgical planning and therapy response assessment for HCC patients.