ESTRO 2024 - Abstract Book

S4499

Physics - Machine learning models and clinical applications

ESTRO 2024

variability. To overcome these, we aimed to develop a deep learning-based framework to automatically predict the achievable dose distribution based on the CT images and delineated structures.

Material/Methods:

We utilized CT images, delineated structures (target volume and organs-at-risk; OAR), and dose distributions of 795 patients who underwent radiotherapy treatments with volumetric modulated arc therapy (VMAT; left breast n = 249, right breast n = 289, left breast with regional lymph nodes n = 124, and right breast with regional lymph nodes n = 133). The breast treatments were conducted with two tangential VMAT arcs (~50° per arc) and the breast with regional lymph nodes with a single, continuous partial VMAT arc (~240°). The original treatment plans were done using the Monaco treatment planning system (Elekta AB, Sweden). MIM Maestro software (MIM software Inc., USA) was used for extracting the data and generating image slices masked with the delineated structures with similar dimensions as in the original CT. A deep learning model was separately trained for each of the four cases, and the patients were further randomly split into training (70%), validation (10%), and test (20%) sets. The CT image together with the masked image were used as an input to the model. The output of the model was the predicted dose distribution for each slice. The model was a modified U-net. The model was trained using the slice-by-slice mean absolute error (MAE) as the loss function. We used Python version 3.10.6 with Tensorflow version 2.10.0 for building the deep learning model. The model performance was assessed by calculating the average slice-by-slice MAE between the original and predicted dose distributions. Moreover, the DVH-based metrics over the OARs were assessed.

Results:

The model achieved a slice-by-slice MAE of 0.12 Gy, 0.14 Gy, 0.18 Gy, and 0.17 Gy for the left breast, right breast, left breast with regional lymph nodes, and right breast with regional lymph nodes cases, respectively. The DVH metrics showed good agreement between the predicted and planned doses for the target volume and organs at risk. Mean absolute errors of OAR mean doses are presented in Table 1 and DVHs with median accuracy are presented in Figure 1.

Table 1. Mean absolute errors ± standard deviation of organ-at-risk mean doses between clinical dose and deep learning-based prediction.

Target

Heart

Ipsilateral lung

Contralateral lung

Contralateral breast

Left breast

0.24 ± 0.28

0.74 ± 0.47

0.10 ± 0.06

0.19 ± 0.13

Right breast

0.11 ± 0.08

0.55 ± 0.39

0.08 ± 0.07

0.14 ± 0.12

Left breast + lymph nodes Right breast + lymph nodes

0.62 ± 0.83

0.82 ± 0.71

0.21 ± 0.18

0.67 ± 0.51

0.31 ± 0.30

0.60 ± 0.32

0.22 ± 0.23

0.59 ± 0.40