Three-dimensional dose and LETD prediction in proton therapy using artificial neural networks

Abstract

Purpose: Challenges in proton therapy include identifying patients most likely to benefit; ensuring consistent, high-quality plans as its adoption becomes more widespread; and recognizing biological uncertainties that may be related to increased relative biologic effectiveness driven by linear energy transfer (LET). Knowledge-based planning (KBP) is a domain that may help to address all three.

Methods: Artificial neural networks were trained using 117 unique treatment plans and associated dose and dose-weighted LET (LET_D ) distributions. The data set was split into training (n = 82), validation (n = 17), and test (n = 18) sets. Model performance was evaluated on the test set using dose- and LET_D -volume metrics in the clinical target volume (CTV) and nearby organs at risk and Dice similarity coefficients (DSC) comparing predicted and planned isodose lines at 50%, 75%, and 95% of the prescription dose.

Results: Dose-volume metrics significantly differed (α = 0.05) between predicted and planned dose distributions in only one dose-volume metric, D_2% to the CTV. The maximum observed root mean square (RMS) difference between corresponding metrics was 4.3 Gy_RBE (8% of prescription) for D_1cc to optic chiasm. DSC were 0.90, 0.93, and 0.88 for the 50%, 75%, and 95% isodose lines, respectively. LET_D -volume metrics significantly differed in all but one metric, L_0.1cc of the brainstem. The maximum observed difference in RMS differences for LET_D metrics was 1.0 keV/μm for L_0.1cc to brainstem.

Conclusions: We have devised the first three-dimensional dose and LET_D -prediction model for cranial proton radiation therapy has been developed. Dose accuracy compared favorably with that of previously published models in other treatment sites. The agreement in LET_D supports future investigations with biological doses in mind to enable the full potential of KBP in proton therapy.

Keywords: artificial intelligence; knowledge-based planning; neural networks; proton therapy.

Figure 1:

Loss curves for neural network models trained to predict radiation dose (top) and dose-weighted linear energy transfer (LET_D) (bottom) are shown.

Figure 2.

An axial slice of a representative test patient is shown with the treatment planning dose (column A) and the neural network–predicted dose (column B). The top row shows the relation of the target (red) to other nearby organs at risk. The center row is zoomed in to show the 50% (green), 75% (yellow), and 95% (red) isodose lines. The bottom row displays the dose in color wash. The prescription dose is 54 Gy_RBE.

Figure 3.

An axial slice of a representative test patient is shown with the Monte Carlo–calculated dose-weighted linear energy transfer (LET_D) (column A) and the neural network–predicted LET_D (column B). The top row shows the relation of the target (red) to other nearby organs at risk (OARs). The center row is zoomed in to show the LET_D distribution around the target and adjacent OARs. The bottom row displays the LET_D as iso-LET_D contours at the 7, 6, 5, 4, and 3 keV/μm levels.

Figure 4.

Violin plots showing the difference in artificial neural network–predicted dose and treatment planning dose for several dose-volume metrics of nearby organs at risk (top – a). The violins represent the distribution of metric differences, with the center circles corresponding to the median difference. Values greater than zero correspond to an overprediction by the neural network. The data are taken from the 18-patient test set. Violin plots showing the difference in artificial neural network-predicted LET_D and Monte Carlo calculated LET_D are shown at the bottom (b).