MS-FANet: Multi-scale feature attention network for liver tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate liver tumor segmentation is crucial for early liver cancer diagnosis.
• Existing segmentation networks struggle with the variable volumes of liver tumors in CT images.
• A need exists for advanced methods to improve segmentation accuracy.

Purpose of the Study:

• To propose a novel multi-scale feature attention network (MS-FANet) for enhanced liver tumor segmentation.
• To address the limitations of current networks in handling diverse tumor sizes.
• To improve the accuracy of liver tumor segmentation in computed tomography (CT) scans.

Main Methods:

• Developed a multi-scale feature attention network (MS-FANet) incorporating a novel residual attention (RA) block and multi-scale atrous downsampling (MAD) in the encoder.
• Integrated a dual-path feature (DF) filter and dense upsampling (DU) in the feature reduction process.
• Evaluated MS-FANet on the public LiTS and 3DIRCADb datasets.

Main Results:

• MS-FANet achieved an average Dice score of 74.2% on the LiTS dataset.
• MS-FANet achieved an average Dice score of 78.0% on the 3DIRCADb dataset.
• The proposed network outperformed most state-of-the-art segmentation networks.

Conclusions:

• MS-FANet demonstrates excellent performance in liver tumor segmentation.
• The network effectively learns features at different scales, adapting to variable tumor volumes.
• The proposed architecture shows significant potential for improving early liver cancer diagnosis through accurate segmentation.