Why Use Position Features in Liver Segmentation Performed by Convolutional Neural Network.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Liver volumetry is crucial in clinical practice, often relying on Computed Tomography (CT) for volume calculation.
• Traditional automatic segmentation algorithms struggle with accuracy, leading to leakage into adjacent organs like the heart and spleen.
• Convolutional Neural Networks (CNNs) show promise for medical image segmentation but require robust input features.

Purpose of the Study:

• To enhance the accuracy of automatic liver segmentation in CT images.
• To address the limitations of existing segmentation algorithms regarding leakage and precision.
• To improve the reliability of liver volumetry for clinical applications.

Main Methods:

• Utilized CNNs, specifically the U-Net architecture, for liver segmentation.
• Incorporated signed distance fields derived from robust anatomical structures (spine, body surface, sagittal plane) as additional input channels.
• Trained and tested the model on two public CT image datasets.

Main Results:

• The proposed method, leveraging additional positional information, demonstrated improved liver segmentation results.
• Quantitative evaluation using Accuracy, Hausdorff distance, and Dice coefficient confirmed the enhanced performance.
• The approach effectively reduced segmentation leakage into surrounding tissues.

Conclusions:

• The integration of anatomical landmark-derived positional information significantly boosts the performance of CNN-based liver segmentation.
• This enhanced segmentation accuracy contributes to more reliable liver volumetry in clinical settings.
• The publicly available code facilitates further research and application of this technique.