Spatial feature fusion convolutional network for liver and liver tumor segmentation from CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Manual segmentation of liver and tumors from CT images is time-consuming and subjective.
• Automated segmentation methods are crucial for improving efficiency and accuracy in radiological analysis.
• Existing methods face challenges due to variations in shape, volume, and image intensity.

Purpose of the Study:

• To develop and evaluate a Spatial Feature Fusion Convolutional Network (SFF-Net) for automatic segmentation of liver and liver tumors in CT images.
• To improve upon the accuracy and efficiency of current liver and liver tumor segmentation techniques.
• To provide a reliable tool for radiologists in treatment planning and decision-making.

Main Methods:

• The SFF-Net utilizes side-outputs from convolutional blocks to leverage multiscale features.
• Skip-connections are incorporated during down-sampling to preserve spatial information.
• Feature fusion blocks (FFBs) merge spatial and semantic features from different network layers.
• A 3D conditional random field (CRF) is applied for final segmentation refinement.

Main Results:

• The SFF-Net was evaluated on the MICCAI 2017 Liver Tumor Segmentation (LiTS) dataset.
• For liver segmentation, achieved Dice Global (DG) score of 0.955 and Dice per case (DC) score of 0.937.
• For tumor segmentation, achieved DG score of 0.746 and DC score of 0.592, with a tumor precision score of 0.369.

Conclusions:

• The SFF-Net effectively learns spatial information through skip-connections and feature fusion.
• The proposed method demonstrates high accuracy in segmenting both the liver and liver tumors from CT images.
• This automated approach has the potential to significantly assist radiologists in clinical practice.