Automatic multi-organ segmentation in computed tomography images using hierarchical convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiation Oncology

Background:

• Accurate segmentation of organs in computed tomography (CT) images is critical for radiation therapy (RT) planning.
• Low soft tissue contrast in CT images presents a significant challenge for precise organ segmentation.

Purpose of the Study:

• To develop and validate a novel two-step hierarchical convolutional neural network (CNN) for automated, accurate segmentation of multiple organs in CT images for RT planning.
• To enhance segmentation accuracy and computational efficiency compared to existing methods.

Main Methods:

• A two-step hierarchical CNN strategy was employed, starting with a coarse segmentation using UNet to define organ-specific regions of interest (ROIs).
• A generative adversarial network (GAN), with a UNet generator and a fully convolutional network discriminator, was used for detailed, fine segmentation within the ROIs.
• The method was validated on male pelvic CTs (prostate, bladder, rectum) and head and neck (H&N) CTs (parotid and submandibular glands).

Main Results:

• The proposed method achieved high segmentation accuracy, with average Dice similarity coefficients (DSC) of 0.91 (prostate), 0.96 (bladder), 0.93 (rectum), 0.88 (parotid glands), and 0.85 (submandibular glands).
• Segmentation of each CT image was performed rapidly, taking an average of 1.4 seconds.
• The H&N segmentation network demonstrated comparable performance on a public dataset.

Conclusions:

• The developed hierarchical CNN with GAN integration provides a fast, accurate, and reproducible solution for segmenting multiple organs in CT images.
• The method shows significant potential for application across various disease sites in radiation therapy planning.