Author: Jing Cai, Zhi Chen, Hong Ge, Yu-Hua Huang, Bing Li, Zihan Li, Ge Ren 👨🔬
Affiliation: Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Department of Radiation Oncology, The Affiliated Cancer Hospital of Zhengzhou University and Henan Cancer Hospital 🌍
Purpose: Algorithms based on subregional respiratory dynamics (SRD) capture spatiotemporal heterogeneity in the ventilation process, though rely on empirical modelings to map surrogate ventilation. Given that avoidance of normal lung tissue during radiotherapy reduces radiation-induced injury, this study investigated the feasibility of a machine learning approach for temporally analyzing SRD extracted from four-dimensional computed tomography (4DCT) scans, utilizing a dual-path recurrent neural network (DPRNN) that integrates local and global cyclical respiratory patterns for function assessment.
Methods: 46 lung cancer patients with 4DCT and nuclear medicine-based ventilation reference (Vref) were retrospectively collected from the VAMPIRE challenge. Lung parenchyma was partitioned into anatomically constrained subregions on the end-expiratory phase image. Respiratory dynamics were characterized through SRD extraction, capturing both local subregional changes and global whole-lung patterns across breathing phases. DPRNN was designed to classify subregions as normal or defective functions, with separate paths processing intensity and volume changes. Data augmentation through cyclic shifting was implemented to capture phase-invariant features in the respiratory cycle. The ventilation distribution from DPRNN (VDPRNN), generated by interpolating subregion-wise classification probabilities, and empirical SRD-based ventilation maps (VESRD) were evaluated against Vref through voxel-wise correlation analysis and functional region overlap metrics.
Results: DPRNN showed promising performance in distinguishing functional from defective lung subregions, with the area-under-the-receiver-operating-characteristic-curve of 0.709 and accuracy of 0.727 for training, and 0.695 and 0.696 for testing, respectively. The generated VDPRNN maps showed significantly improved correlations with Vref scans (Spearman =0.61±0.11, p=0.0004 vs VESRD) and enhanced spatial concordance of high-functioning and low-functioning regions (Dice coefficient=0.62±0.05 and 0.63±0.07, p=0.0277 and 0.0054 vs VESRD).
Conclusion: The study demonstrates the feasibility of using machine learning to analyze SRD patterns for lung function assessment, offering an alternative perspective on leveraging respiratory temporal information for ventilation mapping, with potential applications in improving radiotherapy planning through accurate identification of functional lung regions.