ESTRO 2024 - Abstract Book

S3512

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

844

Digital Poster

Head-to-head comparison for breast RT planning: knowledge-based planning vs. deep learning solution

Daniel Portik 1 , Enrico Clementel 2 , Jérôme Krayenbühl 3 , Nienke Bakx 4 , Nicolaus Andratschke 3 , Coen Hurkmans 4,5

1 European Organisation for Research and Treatment of Cancer (EORTC) Headquarters, Medical Department, Brussels, Belgium. 2 European Organisation for Research and Treatment of Cancer (EORTC) Headquarters, Medical Department, Brussesl, Belgium. 3 University Hospital Zürich, Department of Radiation Oncology, Zürich, Switzerland. 4 Catharina Hospital Eindhoven, Department of Radiation Oncology, Eindhoven, Netherlands. 5 Technical University Eindhoven, Department of Applied Physics and Department of Electrical Engineering, Eindhoven, Netherlands

Purpose/Objective:

To improve efficiency and plan quality, dose prediction methods have been used to automate the planning process. To this end knowledge-based planning (KBP) and deep learning (DL) solutions have been developed. Here we offer the first comparison using the same training and validations sets between these approaches.

Material/Methods:

Two KBP models (Clean and Non-Clean) were created using 90 RT plans, with 15 radiotherapy plans set aside for testing. Results were compared with a DL U-net model and manually produced RT plans for the same dataset. Dutch consensus criteria were used for evaluation. The terms of ‘calculated’ refer to the deliverable plans for both KBP models, with their DL counterparts represented by the ‘mimicked’ term. ‘Estimated’ values refer to the DVH estimated values for both KBP models, these being the equivalents of the DL model’s ‘predicted’ doses. Statistical analysis was done with Wilcoxon-signed rank test and pairwise comparisons with 0.05 significance on estimated, calculated, predicted, and mimicked values.

Results:

Both KBP models equally underestimated mean heart and lung dose (0.5Gy± 0.2Gy and 1.6Gy± 0.5Gy) compared to the clinical plans (1.0Gy±0.4Gy and 1.9Gy± 0.5Gy). In the final calculations the mean lung dose was higher (2.2Gy± 0.5Gy) for both KPB models. The U-Net model resulted in a PTV mean dose of 40.8Gy±0.3Gy, slightly higher than the clinical plans (40.5Gy± 0.2Gy), while both KBP models produced values of 40.6Gy±0.5Gy. PTV D2% values for both KBP models were lower (41.8Gy±0.5Gy) than clinical values (42.1Gy±0.3Gy) and U-net mimicked values (42.4Gy±0.4Gy). PTV coverages were within Dutch consensus criteria. Amongst our KBP models, differences were statistically non-significant for the analysed DVH parameters (Fig 1-2).

PTV

SD

Difference from prescribed dose (%)

Heart (Gy)

SD

Lung (Gy)

SD

Models

(Gy)

(Gy)

(Gy)

(Gy)

Mean

40.5

0.2

1.2

1.0

0.4

1.9

0.5

Clinical

D2%

42.1

0.3

5.0