Deep learning-based 4D robust optimization of intensity—modulated proton therapy for lung cancer radiotherapy

<i>Objective.</i>Intensity-modulated proton therapy (IMPT) provides steep dose gradients but is vulnerable to range uncertainties and respiratory motion, leading to interplay effects in lung cancer radiotherapy. This study aimed to develop a deep learning-based 4D (DL4D) optimization framework to mitigate these challenges. The proposed workflow integrates a DL4D optimization framework combining dose prediction on 4DCTs, water-equivalent thickness variation-guided beam selection (ΔWET-guided beam selection), and dose mimicking to generate 4D-robust IMPT plans.<i>Approach.</i>The planning process uses a U-Net model to predict robust dose distributions based on multiple CT inputs, followed by dose mimicking for plan generation. In this study, data from 62 patients with lung cancer, including 4DCT were used, with dose data generated from the beam angles which were selected based on the ΔWET at different phases. The dataset was split into 42 training, 10 validation, and 10 testing cases. The dose-volume histogram and robustness of the plans were evaluated.<i>Main results.</i>We demonstrated that DL4D plans maintained target coverage across respiratory phases and improved conformity over the robust plans produced by the internal gross tumor volume-override (IGTV-override) CT. The conformity index was higher for DL4D plans both for the IGTV on IGTV-override CT (80.8% vs 69.4%,<i>p</i>= 0.002) and for the GTV in dose accumulation (62.5% vs 55.5%,<i>p</i>= 0.002). Accumulated<i>D</i>_98%of the GTV was close to the prescription dose for both (70.3 Gy vs 71.5 Gy,<i>p</i>= 0.131). OAR doses were clinically comparable.<i>Significance.</i>DL4D optimization with ΔWET-guided beam selection and dose mimicking yields IMPT plans with better conformity, offering an efficient alternative to conventional 4D robust optimization for lung cancer treatment plans.