Automatic computed tomography image segmentation method for liver tumor based on a modified tokenized multilayer...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate liver and tumor segmentation is crucial for diagnosing liver diseases.
• Current deep neural networks (CNNs, ViT) prioritize accuracy over speed, limiting rapid clinical application.
• There is a need for faster segmentation methods without compromising diagnostic accuracy.

Purpose of the Study:

• To develop an automatic computed tomography (CT) image segmentation method for liver tumors.
• The goal is to reduce inference time while maintaining segmentation accuracy.
• To validate the proposed method rigorously through experimental studies.

Main Methods:

• A U-shaped network enhanced with a multiscale attention module and attention gates was developed.
• Modified tokenized multilayer perceptron (MLP) blocks reduce feature dimensions and computational complexity.
• Attention gates and a multiscale attention mechanism improve focus on relevant features and adapt to varying tumor sizes.

Main Results:

• The proposed network achieved a Dice score of 0.713 and an average inference time of 26 ms for liver tumor segmentation.
• Performance metrics include volumetric overlap error (0.39) and average symmetric surface distance (2.04 mm).
• The model was validated on the Liver Tumor Segmentation 2017 (LiTS17) public dataset.

Conclusions:

• The developed network provides efficient liver tumor segmentation with reduced inference time.
• This advancement supports the use of neural networks for rapid clinical diagnosis and treatment planning.
• The findings highlight the potential for faster, AI-driven medical image analysis.