Complete abdomen and pelvis segmentation using U-net variant architecture.

Area of Science:

• Medical imaging analysis
• Radiotherapy treatment planning
• Artificial intelligence in medicine

Background:

• Organ segmentation in computed tomography (CT) is crucial for radiotherapy, but manual segmentation of organs-at-risk is time-consuming.
• Accurate segmentation of all abdominal tissues, including overlooked structures, is necessary for comprehensive treatment planning.

Purpose of the Study:

• To develop and evaluate a fully automated 3D U-Net convolutional neural network (CNN) for segmenting 33 unique organ and tissue structures in the whole abdomen and pelvis from CT images.
• To assess the model's capability in segmenting complex tissues like adipose tissue, skeletal muscle, connective tissue, and vessels.
• To enable quantification of radiation exposure to all abdominal tissues beyond just a few organs-at-risk.

Main Methods:

• A 3D U-Net variant architecture with residual blocks was employed for automated segmentation.
• The model was trained and validated on 66 CT examinations and tested on 18 CT examinations.
• Segmentation accuracy was quantified using the Dice coefficient and compared to human segmentation variability and other state-of-the-art models.

Main Results:

• The 3D U-Net model achieved high accuracy, with Dice coefficients of 0.95 for the liver and 0.93 for the kidneys.
• Accuracies for other organs included 0.79 for the pancreas and 0.69 for the adrenals.
• Model performance was within 5% of human segmentation for 8 out of 19 organs and within 10% for 13 out of 19 organs.

Conclusions:

• The developed CNN demonstrates accuracy comparable to human segmentation, with superior consistency for complex organs.
• Fully automated deep learning-based segmentation of the CT abdomen offers significant potential to enhance both the speed and accuracy of radiotherapy dose prediction for organs-at-risk.