Automatic Contour Refinement for Deep Learning Auto-segmentation of Complex Organs in MRI-guided Adaptive Radiation...

Area of Science:

• Medical Imaging
• Radiation Oncology
• Computational Anatomy

Background:

• Accurate auto-segmentation of abdominal structures in daily MRI is critical for adaptive radiation therapy (ART).
• Current deep learning auto-segmentation methods struggle with complex abdominal anatomy, necessitating contour correction.
• Manual contour correction is time-consuming and may delay ART treatment delivery.

Purpose of the Study:

• To develop and evaluate an automatic contour refinement (ACR) process for correcting inaccurate auto-segmented contours in abdominal MRI.
• To improve the accuracy and efficiency of auto-segmentation for MRI-guided adaptive radiation therapy (ART).

Main Methods:

• An improved level set-based active contour model (ACM) was implemented for the ACR process.
• The ACR process was tested on 80 abdominal MRI datasets with deep learning-based auto-segmentation.
• Contour accuracy was evaluated using Dice Similarity Coefficient (DSC), Mean Distance to Agreement (MDA), Surface DSC, and Added Path Length (APL).

Main Results:

• The ACR process corrected a portion (3%-39%) of auto-segmented contours to meet clinical acceptability standards (DSC >0.8, MDA <3 mm).
• For combined bowels with major errors, mean DSC improved from 0.34 to 0.59, and mean MDA decreased from 7.02 mm to 5.23 mm.
• The ACR process achieved clinically acceptable contours for 39% of cases with minor errors and operated in under 2 seconds per subregion on a GPU.

Conclusions:

• The ACR process effectively corrects inaccurate contours from deep learning-based auto-segmentation of complex abdominal anatomy in MRI.
• This ACR method can be integrated into auto-segmentation workflows to accelerate MRI-guided ART.
• The ACR process offers a promising solution for improving the efficiency and accuracy of ART planning.