Click here to

Session: AI in Treatment Planning [Return to Session]

Aperture and Control Point Weight Prediction for Breast VMAT Using Deep Learning

L Vandewinckele1*, T Reynders2, S Petillion2, C Weltens1,2, F Maes3,4, W Crijns1,2, (1) Department of Oncology, Laboratory of Experimental Radiotherapy, KU Leuven, BE-Belgium, (2) Department of Radiation Oncology, UZ Leuven, BE-Belgium, (3) Department ESAT/PSI, KU Leuven, BE-Belgium,(4) Medical Imaging Research Center, UZ Leuven, BE-Belgium,


SU-H300-IePD-F4-5 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 4

Purpose: State-of-the-art research in radiotherapy treatment planning using deep learning is directed towards indirect treatment plan prediction via dose or fluence prediction from patient anatomy after which a conversion to VMAT apertures and MU is needed to obtain the final treatment plan. This work aims to predict VMAT apertures and MU directly from patient anatomy for a three arc VMAT breast plan.

Methods: A dataset consisting of 134 patients (training: 110; testing: 24) treated for right breast cancer was replanned for VMAT using our clinical RapidPlanᵀᴹ model (doses: PTV_breast: 45.57Gy; PTV_boost: 55.86Gy). A CNN predicts from CT and contours, an aperture and MU per control point. The input consists of 2D projections of the CT and contours along the BEV. The network predicts for all control points an image of which the shape equals the MLC aperture and the intensity the MU. The shape and MU are used to build an RTplan dicom. Next this is imported in Eclipse to perform a dose calculation of the predicted plan.

Results: The mean DICE/ASSD between the ground truth (gt) and predicted MLC apertures varies over the control points between 0.20/29.80mm for small apertures (min 29.03cm²) and 0.78/1.23mm for large apertures (max 109.20cm²). The mean absolute error between the gt and predicted MU is 0.84±1.02MU (mean MU gt: 2.08±1.39; pred: 1.33±0.61). After dose calculation, the plans achieved clinically acceptable doses onto the OAR (mean doses: Heart: 1.12±0.30Gy; Contralateral Lung: 0.55±0.17Gy; Ipsilateral Lung: 5.78±1.19Gy; Contralateral Breast: 0.90±0.40Gy; Liver: 1.45±0.64Gy). However, improvement is needed for the PTV doses (D95%: PTV_boost: 47.36±2.70; PTV_breast: 37.92±3.57).

Conclusion: This work shows the feasibility of direct parameter prediction for VMAT. The network manages in sparing the OAR, but PTV coverage needs improvement. This could be achieved by improved prediction of small apertures and higher MUs with more variation.

Funding Support, Disclosures, and Conflict of Interest: Liesbeth Vandewinckele is supported by a Ph.D. fellowship of the research foundation - Flanders (FWO), under grant number 1SA6121N. This work is supported by Varian Medical Systems.


Treatment Planning, Breast


TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email