Liver Extraction Using Residual Convolution Neural Networks From Low-Dose CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate liver segmentation is vital for diagnosing hepatic diseases and planning treatments.
• Low-dose CT (LDCT) reduces radiation exposure but presents challenges like noise and indistinct boundaries.
• Existing methods struggle with the image quality degradation inherent in LDCT scans.

Purpose of the Study:

• To develop an efficient and precise liver extraction method specifically for LDCT images.
• To facilitate surgical planning and post-operative assessment with reduced radiation risk.
• To improve the accuracy of liver modeling in medical imaging.

Main Methods:

• Utilized a residual convolutional neural network (LER-CN) for liver extraction.
• Incorporated noise removal and structure preservation components to refine the model.
• Employed patch-based training and backpropagation gradient descent algorithms.
• Evaluated the method on 150 abdominal CT scans and the MICCAI Sliver07 dataset.

Main Results:

• LER-CN achieved a Dice Similarity Coefficient of up to 96.5%.
• Demonstrated a reduced volumetric overlap error of 4.30%.
• Achieved an average symmetric surface distance of less than 1.4.
• Showed competitive performance against state-of-the-art methods.

Conclusions:

• LER-CN provides an efficient and accurate solution for liver extraction from LDCT images.
• The method supports medical applications requiring high precision with minimized patient radiation.
• This technique aids in better surgical planning and post-operative assessment.