3051 - Deep Learning-Based Dynamic Blood Dose Prediction for Non-Small Cell Lung Cancer

Purpose/Objective(s): Post-treatment Severe Radiation-Induced Lymphopenia (SRIL) is strongly associated with adverse clinical outcomes following radiotherapy (RT), i.e., tumor recurrence and poor overall survival. Accurate quantification of the radiation dose delivered to circulating lymphocytes is therefore essential for assessing SRIL risk. To address this need, HEmatological DOSe (HEDOS) has been developed to estimate dynamic blood dose as a surrogate for the exposure to circulating lymphocytes. Although HEDOS provides reliable dynamic blood dose estimation, it can take up to several minutes to compute the blood dose from dose-volume histograms (DVHs) established by treatment planning and patient-specific clinical parameters. This computational cost limits its practical integration into the RT planning workflow, which relies on iterative optimization of treatment parameters to achieve clinically acceptable DVHs. In this study, we aim to develop a deep learning-based framework for rapid dynamic blood dose estimation methods to integrate blood dose evaluation into RT planning processes.

Materials/Methods: DVHs of organs-at-risk (OARs) and clinical parameters were collected for 93 patients with Non-Small Cell Lung Cancer (NSCLC) via PortPy benchmark dataset. The ground-truth blood dose histograms were calculated using HEDOS. The patients’ data were split into training (n = 57), internal validation (n = 18), and external testing (n = 18) cohorts. We developed a multi-modal deep neural network that accepts the same input as HEDOS: OAR DVHs and clinical parameters (sex, dose rate, blood volume, cardiac output, number of fractions, and beam-on/off time). High-dimensional features were extracted from OAR DVHs using 1D convolutional layers, and a Transformer encoder was utilized to capture features from the clinical parameters. Subsequently, the features were concatenated and processed through fully connected layers to predict blood dose histograms ranging from 0 to 10 Gy at 0.1 Gy intervals. The model was trained to minimize Kullback–Leibler (KL) divergence between predicted and ground-truth histograms. After completing the training, the prediction accuracy was evaluated using KL divergence and D90%, the minimum dose delivered to 90% of blood volume.

Results: The predicted blood dose histograms demonstrated strong agreement with the ground-truth histograms, with a KL divergence of 0.23. The D90% derived from the predicted blood dose histograms showed mean-absolute error of 4.97%, compared to those derived from the ground-truth blood dose histograms. Notably, the proposed framework completed blood dose prediction in 28 ms per patient.

Conclusion: The proposed deep learning-based blood dose prediction framework enables dynamic blood dose estimation at a sub-second level. Consequently, the dynamic blood dose can be incorporated as a clinical objective during RT planning without introducing substantial computational burden, thereby supporting strategies to mitigate SRIL risk.