A novel MRI segmentation method using CNN-based correction network for MRI-guided adaptive radiotherapy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• MRI-guided adaptive radiotherapy (MR-IGART) requires accurate organ segmentation for treatment planning.
• Manual contouring of organs like the liver, kidneys, stomach, bowel, and duodenum is time-consuming.
• Deep learning (DL) offers potential for automating and accelerating this segmentation process.

Purpose of the Study:

• To develop and validate a DL model for accurate segmentation of the liver, kidneys, stomach, bowel, and duodenum in 3D MR images.
• To expedite the contouring process in MR-IGART.
• To improve the efficiency of radiation therapy planning.

Main Methods:

• A novel DL model combining a voxel-wise label prediction CNN with a correction network was proposed.
• The correction network utilized dense blocks and sub-networks to enforce anatomical constraints and refine segmentation accuracy.
• The model was trained and validated on 110 patient datasets and tested on 10, evaluating segmentation using Dice coefficient and Hausdorff distance.

Main Results:

• The DL model achieved high segmentation accuracy for abdominal organs, with Dice coefficients ranging from 65.5% (duodenum) to 95.3% (liver).
• The proposed correction network outperformed traditional conditional random field (CRF) methods.
• Automated segmentation significantly reduced manual contouring time by fourfold compared to manual contouring from scratch.

Conclusions:

• The developed DL method provides accurate automatic segmentation of key abdominal organs in 3D MR images.
• This automated approach substantially expedites manual contouring, making it a valuable tool for MR-IGART.
• The findings support the integration of advanced AI techniques to enhance radiotherapy workflow efficiency and precision.