Deep learning based automatic contour refinement for inaccurate auto-segmentation in MR-guided adaptive radiotherapy.

Area of Science:

• Medical Physics
• Radiation Oncology
• Medical Imaging

Background:

• Accurate auto-segmentation is critical for magnetic resonance-guided adaptive radiation therapy (MRgART).
• Current deep learning auto-segmentation (DLAS) models often produce clinically unacceptable contours, especially for abdominal organs.
• Previous work introduced an active contour model (ACM) for automatic contour refinement (ACR) to partially correct DLAS contours.

Purpose of the Study:

• To develop a deep learning-based ACR (DL-ACR) model to enhance the accuracy of DLAS-generated bowel contours.
• To integrate DL-ACR with an existing ACM-ACR workflow for improved contour refinement.
• To evaluate the combined ACR workflow's effectiveness in improving contour accuracy for MRgART.

Main Methods:

• A DL-ACR model was trained and tested on bowel contours from 160 MR datasets (76 MR-simulation, 84 MR-Linac).
• Contours were classified using AAPM TG-132 recommendations and an in-house model.
• For major-error contours, DL-ACR was applied after ACM-ACR; for minor-error contours, DL-ACR was used directly. The workflow was evaluated on 25 independent image sets.

Main Results:

• Using TG-132 classification on MR-simulation data, sequential ACM-ACR and DL-ACR improved 30% of major-error contours to acceptable.
• DL-ACR alone improved 43% of minor-error contours to acceptable.
• The combined ACR workflow demonstrated substantial improvements in contour accuracy, reducing the need for manual editing.

Conclusions:

• The developed DL-ACR model, when combined with ACM-ACR, significantly enhances the accuracy of DLAS bowel contours.
• This improved contour accuracy is crucial for accelerating the segmentation process in MRgART.
• The workflow effectively minimizes manual contour editing, leading to greater efficiency in clinical practice.