A segmentation method with a large vision model for magnetic resonance imaging-guided adaptive radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Segmentation is a time-intensive step in MRI-guided adaptive radiotherapy (MRIgART).
• Existing models like the Segment Anything Model (SAM) require extensive manual input (clicks, bounding boxes) for medical imaging segmentation.
• This limits the efficiency and applicability of advanced AI in clinical workflows.

Purpose of the Study:

• To introduce SAM-ART, a large vision model designed to enhance MRIgART segmentation accuracy.
• To integrate personalized patient information into the SAM framework for improved performance.
• To reduce the manual interaction required for accurate medical image segmentation.

Main Methods:

• Utilized planning CT (pCT), approved contours, and daily MRI (dMRI) from 38 prostate and 10 rectal cancer patients.
• Developed SAM-ART with an image encoder, prompt encoder, and mask decoder.
• Employed mask and box prompts derived from pCT contours propagated via deformable image registration (DIR) to dMRI, incorporating patient-specific data.

Main Results:

• SAM-ART achieved a mean Dice Similarity Coefficient (DSC) of 0.934 ± 0.023, outperforming DIR (0.873 ± 0.063) and traditional deep learning (tDL) (0.887 ± 0.056).
• Mask/box prompts yielded superior results (DSC 0.934) compared to point (0.910) or box (0.921) prompts alone.
• 89.38% of segmentations met clinical acceptance criteria (DSC ≥ 0.85, Hausdorff distance ≤ 5 mm, Mean Distance to Agreement ≤ 1.5 mm), minimizing manual correction needs.

Conclusions:

• A novel SAM-based segmentation method integrating personalized information and optimized prompts was developed.
• The proposed SAM-ART significantly outperformed baseline methods in segmentation accuracy for MRIgART.
• The method substantially reduces the need for manual contour revisions, streamlining clinical workflows.