Deep learning-based GTV contouring modeling inter- and intra- observer variability in sarcomas.

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Manual gross tumor volume (GTV) delineation for radiation therapy is costly and lacks reproducibility due to inter- and intra-observer variability.
• Existing automated contouring methods often disregard this variability, missing opportunities to leverage confidence information from multiple delineations.

Purpose of the Study:

• To develop an automatic GTV contouring method for soft-tissue sarcomas using deep learning.
• To integrate inter- and intra-observer variability into the learning model for more robust contour prediction.

Main Methods:

• A deep convolutional neural network was trained on CT images from 68 sarcoma patients.
• Inter- and intra-observer variability was quantified by voxel inclusion frequency to create confidence levels.
• The model learned GTV confidence maps by integrating this variability.

Main Results:

• The automated method achieved a Dice score of 87% and a Hausdorff distance of 14 mm when compared to consensus contours.
• Performance was evaluated against confidence maps generated by four expert readers.

Conclusions:

• The proposed deep learning method accurately predicts GTV contours by effectively utilizing observer variability.
• This automated approach has the potential to enhance clinical workflow in radiation therapy planning.