Efficient two-step liver and tumour segmentation on abdominal CT via deep learning and a conditional random field

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual segmentation of liver and tumours in CT scans is laborious and time-consuming.
• Accurate segmentation is crucial for effective hepatic surgical planning.

Purpose of the Study:

• To develop a fully automated two-step method for liver and tumour segmentation from CT scans.
• To improve the accuracy and efficiency of segmentation for surgical planning.

Main Methods:

• A cascade framework employing a fractal residual U-Net (FRA-UNet) for initial liver and tumour segmentation.
• Refinement of tumour segmentation using a 3D conditional random field (CRF).
• Utilizing improved fractal residual (FR) structures and deep residual blocks for enhanced feature extraction.

Main Results:

• Achieved high Dice Similarity Coefficients (DSCs): 97.13% for liver and 71.78% for tumours on the LiTS dataset.
• Obtained DSCs of 97.18% for liver and 68.97% for tumours on the 3DIRCADb dataset.
• Outperformed most existing state-of-the-art segmentation networks.

Conclusions:

• The proposed FRA-UNet combined with 3D CRF offers an effective and automated solution for liver and tumour segmentation.
• The method enhances segmentation accuracy and efficiency, benefiting hepatic surgical planning.
• The novel architecture and refinement technique address limitations of previous methods, reducing oversegmentation.