Automated Contouring of Contrast and Noncontrast Computed Tomography Liver Images With Fully Convolutional Networks

Area of Science:

• Medical imaging
• Radiology
• Computational anatomy

Background:

• Liver deformability complicates focal treatment, challenging rigid registration.
• Biomechanical modeling requires time-intensive manual liver segmentation.

Purpose of the Study:

• To investigate fully convolutional networks for rapid and accurate automated liver segmentation.
• To remove the temporal bottleneck for biomechanical modeling in liver treatments.

Main Methods:

• Trained and evaluated three architectures (VGG-16, DeepLabv3+, 3D UNet) on CT scans from 183 patients and 30 challenge scans.
• Assessed accuracy using Dice similarity coefficient and mean surface distance.
• Performed qualitative evaluation by radiation oncologists on 50 independent cases.

Main Results:

• DeepLabv3+ achieved mean surface distances of 0.99 mm (contrast), 1.12 mm (non-contrast), and 1.48 mm (MICCAI).
• 60% of autosegmentations were preferred to manual contours in blinded comparisons.
• 96% of autosegmentations were deemed clinically acceptable by reviewers.

Conclusions:

• Automated liver segmentation using fully convolutional networks is accurate and efficient.
• This method is preferred over manual segmentation and has potential for clinical integration.
• Rapid, accurate autosegmentation enables efficient biomechanical model-based registration in clinical workflows.