Automated treatment planning for proton pencil beam scanning using deep learning dose prediction and dose-mimicking optimization

Abstract

Purpose: The purpose of this study is to investigate the use of a deep learning architecture for automated treatment planning for proton pencil beam scanning (PBS).

Methods: A 3-dimensional (3D) U-Net model has been implemented in a commercial treatment planning system (TPS) that uses contoured regions of interest (ROI) binary masks as model inputs with a predicted dose distribution as the model output. Predicted dose distributions were converted to deliverable PBS treatment plans using a voxel-wise robust dose mimicking optimization algorithm. This model was leveraged to generate machine learning (ML) optimized plans for patients receiving proton PBS irradiation of the chest wall. Model training was carried out on a retrospective set of 48 previously-treated chest wall patient treatment plans. Model evaluation was carried out by generating ML-optimized plans on a hold-out set of 12 contoured chest wall patient CT datasets from previously treated patients. Clinical goal criteria and gamma analysis were used to compare dose distributions of the ML-optimized plans against the clinically approved plans across the test patients.

Results: Statistical analysis of mean clinical goal criteria indicates that compared to the clinical plans, the ML optimization workflow generated robust plans with similar dose to the heart, lungs, and esophagus while achieving superior dosimetric coverage to the PTV chest wall (clinical mean V95 = 97.6% vs. ML mean V95 = 99.1%, p < 0.001) across the 12 test patients.

Conclusions: ML-based automated treatment plan optimization using the 3D U-Net model can generate treatment plans of similar clinical quality compared to human-driven optimization.

Keywords: deep learning; proton therapy.

FIGURE 1

Workflow for traditional treatment planning (a) and ML‐based treatment planning (b).

FIGURE 2

CT scan slices of a patient receiving PBS irradiation of the chest wall with the following contoured targets and OARs: heart (purple), PTV chest wall (red), esophagus (green), lungs (blue), and PTV nodes which includes the axilla, supraclavicular lymph nodes and internal mammary chain (yellow).

FIGURE 3

Architecture of the 3D U‐Net model. Blue boxes represent 3‐dimensional feature volumes, with the number of channels indicated by the top number. Max pooling was implemented with a pool size and stride of 2, thereby halving the spatial dimensions when applied. All convolutional layers, except the last, were followed by a ReLU activation function. The total number of parameters in the model was 1 194 857. Concatenated layers are indicated by a lighter blue color.

FIGURE 4

Loss curves for the training of the U‐Net model. The epoch chosen for postprocessing, corresponding to the lowest validation loss value, is highlighted by the dashed line.

FIGURE 5

Axial CT slice with distributions of the clinical, ML‐predicted, and ML‐optimized treatment plans for a chest wall target (red contour) (upper) and corresponding target and OAR DVH curves (lower).

FIGURE 6

Box plots across all 12 test patients for OARs and treatment targets for the clinical plans (blue) and ML‐optimized plans (orange). Median clinical goal values are illustrated by the solid black line within each box‐whisker plot and clinical goal thresholds are illustrated with the black dotted lines.

FIGURE 7

Box plot comparison of clinical goal percent differences between clinical and ML‐optimized plans for target volumes (a) and OARs (b). Dotted lines represent 0% difference between the clinical and ML‐optimized plans.

FIGURE 8

DVH curves for the CTV chest wall (red), total lung (blue), esophagus (green), and heart (pink) for perturbed dose scenarios across all 12 permutations of setup (5 mm) and range (3%) uncertainty for a ML‐optimized chest wall plan for one test patient. Dashed lines represent the DVH curve for the nominal plan.

FIGURE 9

Dose distributions for the clinical (upper left) and the ML‐optimized (upper right) plan and the histogram for the 3D gamma comparison of the two (bottom left) as well as a 2D gamma value map (bottom right) for one test patient.

FIGURE 10

Dose profile measured with an ion chamber array (a), corresponding TPS‐calculated 2D dose profile (b), and 2D gamma value map (c) using 3%/3 mm pass fail criteria for a test ML‐optimized chest wall patient.