Automatic plan selection using deep network-A prostate study

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Healthcare

Background:

• High-dose-rate (HDR) brachytherapy planning benefits from multicriteria optimization (MCO) algorithms.
• MCO generates numerous Pareto optimal plans rapidly, shifting focus to selecting the best plan from thousands.

Purpose of the Study:

• Introduce novel, visual-like criteria beyond traditional dose-volume histogram (DVH) metrics for plan evaluation.
• Develop and implement a deep learning (DL) framework for automatic selection of optimal HDR brachytherapy plans.

Main Methods:

• Train a DL algorithm using new visual-like criteria (bladder, rectum, urethra, prostate cold spot) and standard DVH metrics.
• Input for the DL model includes 3D dose and anatomical mask images for plan ranking and selection.
• Algorithm trained on 835 prostate cancer patients and validated on 20 patients previously assessed by clinical medical physicists.

Main Results:

• The DL network ranks 2000 plans in 10 seconds, significantly faster than expert manual selection.
• Four DL networks were trained, offering trade-offs between target coverage and organs at risk (OAR) sparing.
• The best DL network's plan selection showed no statistical difference compared to expert choices for multiple criteria.

Conclusions:

• The DL-based approach is flexible, allowing custom criteria and trade-offs for plan quality.
• This fast and robust method adds minimal time to MCO planning, showing strong potential for clinical adoption.