Automatic liver tumor segmentation used the cascade multi-scale attention architecture method based on 3D U-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate liver tumor segmentation in CT images is challenging due to complex tumor structures and low contrast with surrounding tissues.
• Existing methods often struggle to effectively capture multi-scale information and contextual details necessary for precise segmentation.

Purpose of the Study:

• To propose an end-to-end deep learning method for accurate liver tumor segmentation from CT images.
• To enhance feature extraction and fusion capabilities for improved segmentation performance.

Main Methods:

• A cascade network structure incorporating a Side-output Feature Fusion Attention block for multi-level feature fusion and attention-guided information focus.
• Utilizing an Atrous Spatial Pyramid Pooling Attention block to extract multi-scale semantic features.
• Employing a Multi-scale Prediction Fusion block for comprehensive feature integration across network layers.

Main Results:

• The proposed method achieved a Dice per Case of 0.665 on the LiTS dataset and 0.719 on the 3DIRCADb dataset.
• Global Dice scores of 0.812 (LiTS) and 0.784 (3DIRCADb) were obtained, demonstrating robust performance.
• Evaluation confirmed the effectiveness of individual modules within the proposed segmentation framework.

Conclusions:

• The developed method significantly outperforms the baseline 3D U-Net and other U-Net variant-based approaches for liver tumor segmentation.
• The proposed approach offers a superior solution for the challenging task of liver tumor segmentation in medical imaging.