Esophagus segmentation in CT via 3D fully convolutional neural network and random walk

Area of Science:

• Medical Imaging
• Radiotherapy
• Computational Anatomy

Background:

• Precise delineation of organs at risk is critical in radiotherapy planning.
• Automated segmentation methods show promise but struggle with organs like the esophagus due to shape and contrast.
• Manual esophagus segmentation is time-consuming and prone to errors.

Purpose of the Study:

• To develop an automated method for esophagus segmentation from CT images.
• To improve the accuracy and efficiency of organ delineation for radiotherapy treatment planning.

Main Methods:

• A 3D fully convolutional neural network (CNN) generates a probability map.
• An active contour model (ACM) refines the initial segmentation.
• A random walker algorithm, incorporating CNN outputs and CT Hounsfield values, performs the final segmentation.

Main Results:

• The proposed algorithm achieved a mean Dice coefficient of 0.76 ± 0.11.
• Average symmetric square distance was 1.36 ± 0.90 mm, and average Hausdorff distance was 11.68 ± 6.80 mm.
• The method outperformed existing approaches on the Synapse dataset, indicating state-of-the-art performance.

Conclusions:

• The combined CNN and random walker approach accurately segments the esophagus.
• 3D convolutions leverage spatial context for efficient, volume-wise predictions.
• The fully automatic process provides delineations in strong agreement with the gold standard.