A lightweight neural network with multiscale feature enhancement for liver CT segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate segmentation of abdominal Computed Tomography (CT) scans is crucial for diagnosing and treating visceral organ diseases like hepatocellular carcinoma.
• Existing methods may require significant computational resources or lack precision.

Purpose of the Study:

• To propose a novel neural network, Res-PAC-UNet, for precise and efficient liver CT segmentation.
• To offer a low disk utilization method for medical image analysis.

Main Methods:

• Developed a novel neural network (Res-PAC-UNet) integrating a fixed-width residual UNet backbone with Pyramid Atrous Convolutions.
• Trained the network on the medical segmentation decathlon dataset.
• Employed a modified surface loss function for training.

Main Results:

• The Res16-PAC-UNet achieved a Dice coefficient of 0.950 ± 0.019 with fewer than 0.5 million parameters.
• The Res32-PAC-UNet achieved a Dice coefficient of 0.958 ± 0.015 with approximately 1.2 million parameters.
• Demonstrated high quantitative and qualitative performance in liver CT segmentation.

Conclusions:

• Res-PAC-UNet offers a precise and computationally efficient solution for liver CT segmentation.
• The proposed network is suitable for clinical applications requiring accurate analysis of abdominal CT scans.
• This approach contributes to improved diagnostic capabilities for visceral organ diseases.