ESTRO 2024 - Abstract Book

S4505

Physics - Machine learning models and clinical applications

ESTRO 2024

Material/Methods:

We utilized a previously developed deep learning-based dose prediction model for breast and regional lymph node irradiation based on a modified U-net architecture. The model was integrated into clinical practice by implementing the dose prediction in MIM Maestro software (version 7.3.3, MIM software inc., USA) with a custom-made Python (version 3.7.9) extension. The DICOM images (dose planning CT and delineated structures) were automatically fed into the pipeline after delineation of the structures and the predicted dose distribution (prescription dose 40.05 Gy) was returned to MIM. Moreover, this prediction was used to derive dose-volume histograms (DVHs) which in turn were used to derive dose constraints into the Monaco treatment planning system (version 6.1.3.0, Elekta AB, Sweden) through automatic template generation. Subsequently, an automated Monaco script was developed to automatize the optimization of the generated template plans. Three iterations of fluence optimization were allowed. After each iteration, PTV-skin (PTV cropped by 5 mm from body contour) dose coverage was evaluated. If the predefined limit was not reached (PTV-skin D40Gy > 97%) two constraints with the highest weights were relaxed by 10% and fluence was reoptimized. When the desired dose coverage was reached, or after three iterations of fluence optimization, segment shape optimization was performed. The final dose distribution was calculated using the Monte Carlo algorithm with statistical uncertainty of 1%. All the treatment plans were normalized to the mean of the PTV-skin dose. The generalizability of the clinical implementation was verified by utilizing data (CT and structure set) from another RT center (left breast + regional lymph nodes, n = 49 patients treated with VMAT technique) that was not used for training the model.

Results:

Deep learning-based automatic treatment planning can successfully be implemented within the normal clinical workflow. The results showed that the plans generated using the automated planning corresponded well with the deep learning-based dose predictions (Figures 1 and 2). Moreover, the results illustrate that the automated planning approach can be generalized to data acquired at another RT center.