Purpose: Understanding physician’s preference in treatment planning is necessary to develop automatic treatment planning approaches capable of generating plans favorable by the physician. The purpose of this study is to develop a deep learning based Virtual Physician Network (VPN) that can understand physician’s preference on plan approval for prostate cancer Stereotactic Body Radiation Therapy (SBRT) plans and suggest where to improve, if a plan is rejected.
Methods: The VPN takes one Planning Target Volume (PTV) and eight Organs at Risk (OAR) contours, as well as a dose distribution of a plan seeking approval as input. It outputs the probability of approving the plan, and a dose distribution indicating improvements to be applied to the input dose. VPN includes two sub-networks. VPN1 uses an attention-gated U-Net to predict where to improve in the input dose. VPN2 has the first half architecture of VPN1 followed by fully-connected layers to output a probability of having the given plan approved. Due to the lack of unapproved plans in our database, two networks are alternatively trained under an adversarial framework. The intermediate results generated by VPN1 in the training process are suboptimal dose, which can naturally serve as training data for VPN2. 44 prostate cancer patients who received 45 Gy in 5-fraction SBRT were used in 4-fold cross-validation.
Results: VPN can differentiate high- and low- quality plans with Area under the curve 0.91±0.02. For unapproved plans, after applying VPN’s suggestion on dose improvement, V50 of bladder, rectum, and penile bulb, max dose of urethra, V95 and D95 of PTV agreed with ground truth with differences of 0.59±0.78%, 0.8±1.01%, 0.04±0.03%, 0.75±0.46Gy, 0.34±0.21%, and 0.75±1.27Gy.
Conclusion: VPN was developed to accurately predict a physician’s preference in plan approval and can provide useful suggestions on how to improve the dose distribution.
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation