Transfer learning for DL-based Synthetic CT after reconstruction algorithm upgrade in a proton therapy clinic

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) models generate synthetic CT (sCT) images for adaptive proton therapy.
• DL models are vulnerable to data drift from CBCT software upgrades, impacting image quality.

Purpose of the Study:

• Assess transfer learning for adapting sCT models to new CBCT image quality.
• Evaluate the sustainability of sCT workflows post-CBCT software upgrades.

Main Methods:

• Re-trained DCNN and cycleGAN sCT models using transfer learning on 69 head and neck cancer patients.
• Trained transfer learning models (TL-models) with 5-40 image pairs; benchmarked against a full post-upgrade model.
• Evaluated sCT image quality and dosimetric accuracy using clinical plans and NTCP models.

Main Results:

• Transfer learning models showed comparable Mean Absolute Error (MAE) to the full post-upgrade model.
• Dosimetric accuracy remained high, with minimal mean dose differences (-0.98% to 3.15%).
• Normal tissue complication probability (NTCP) deviations were negligible (-0.03% to 0.0027%).

Conclusions:

• Transfer learning strategies effectively minimize disruptions in sCT-dependent workflows.
• Small data requirements for transfer learning reduce downtime and data acquisition needs.