Author: Yang Lei, Haibo Lin, Tian Liu, Charles B. Simone, Shouyi Wei, Ajay Zheng 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, New York Proton Center 🌍
Purpose: Radiation-induced lung injury (RILI), encompassing pneumonitis and fibrosis, represents a critical dose-limiting factor in lung cancer radiation therapy. Variability in treatment outcomes is often attributed to differential prioritization of lung regions based on their functional significance. Sparing highly functional lung areas by leveraging the spatial heterogeneity of lung function could mitigate RILI risk. This study aims to develop a deep learning-based deformable image registration method to derive functional lung data from 4DCT scans, facilitating improved proton therapy dose optimization and reducing RILI.
Methods: We employed a cycle-diffusion network to achieve robust and consistent deformable registration between the inhale and exhale phases of 4DCT scans. The network estimates deformation vector fields that capture significant lung motion, subsequently used to generate ventilation images (VIs). The model was trained and validated using a three-fold cross-validation on the VAMPIRE database, which includes 27 subjects with paired CT scans and reference VIs (RefVIs). Clinical feasibility was demonstrated by applying the model to 12 consecutive lung cancer patients undergoing proton therapy, with registration accuracy and VI fidelity assessed through quantitative metrics.
Results: Registration accuracy on the clinical dataset was evaluated within the whole body using mean absolute error and target registration error, yielding mean values of 46.3±10.6HU and 1.3±0.5mm, respectively. On the VAMPIRE dataset, estimated VIs were compared to RefVIs using the Spearman correlation coefficient (SP) and Dice similarity coefficient (DSC). The SP was 0.43±0.12, whereas DSC scores were 0.65±0.11, 0.73±0.10, and 0.59±0.09 for high, medium, and low ventilation regions, respectively.
Conclusion: A novel cycle-diffusion-based method was developed to estimate VIs from paired CT scans, offering a strategic solution for quantifying respiratory motion and enabling functional lung avoidance in thoracic proton therapy. Evaluation metrics demonstrate the feasibility of the proposed method. By incorporating functional information, this approach could reduce RILI risk and improve patient outcomes.