Progressive auto-segmentation for cone-beam computed tomography-based online adaptive radiotherapy.

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Accurate auto-segmentation of organs-at-risk (OARs) and targets is critical for online adaptive radiotherapy (ART).
• Current auto-segmentation methods using cone-beam computed tomography (CBCT) often lack clinical acceptability.
• Existing approaches fail to utilize historical data from initial plans and prior adaptive fractions to improve segmentation precision.

Purpose of the Study:

• To develop a novel framework for enhancing CBCT auto-segmentation accuracy in online ART.
• To incorporate temporal context from initial planning and previous adaptive fractions into the segmentation process.
• To refine segmentation precision for current fraction CBCT images by leveraging historical data.

Main Methods:

• Introduction of the LSTM-UNet architecture, integrating Long Short-Term Memory (LSTM) units into U-Net skip connections.
• Utilizing data from a patient's initial plan and previous adaptive fractions to retain temporal information.
• Model pre-training with simulated data followed by fine-tuning on a clinical dataset.

Main Results:

• The proposed LSTM-UNet model achieved an average Dice similarity coefficient of 79% for 8 Head & Neck organs and targets.
• This represents a significant improvement over a baseline model without prior knowledge (52%) and a baseline with prior knowledge but no memory (78%).

Conclusions:

• The LSTM-UNet model effectively utilizes information from prior fractions, outperforming baseline segmentation frameworks.
• This approach reduces the need for manual revision of auto-segmentation results by clinicians.
• The model shows promise for integration into online ART workflows, providing precise segmentation on synthetic CT images.