2410 - Integrating Radiation-Induced Functionally-Weighted Airway Damage Modeling with Radiation-Induced Parenchymal Damage Modeling For Prediction of Post-RT Ventilation Loss

E. M. Vicente¹, N. Gutierrez², J. Oakes³, J. Cammin¹, A. Gopal¹, J. Kipritidis⁴, A. Modiri¹, S. Mossahebi¹, P. Mohindra⁵, W. Citron¹, M. M. Matuszak⁶, R. D. Timmerman⁷, and A. Sawant¹; ¹University of Maryland, School of Medicine, Baltimore, MD, ²CMU, Pittsburgh, PA, ³Northeastern University, Boston, MA, ⁴Northern Sydney Cancer Centre, Royal North Shore Hospital, Sydney, NSW, Australia, ⁵University Hospitals Seidman Cancer Center, Cleveland, OH, ⁶Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, ⁷Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX

Purpose/Objective(s): A common limitation in clinical as well as investigational lung radiotherapy (RT) is that radiation-induced lung injury (RILI) is modeled primarily based on local-regional parenchymal damage. The predictive performance of these models is relatively poor. Mounting evidence from our studies and others suggests that damage to airways and pulmonary vasculature significantly impacts post- RT lung function. We hypothesize that including radiation-induced damage to airways in current (parenchyma-based) predictive dose-response models will improve the estimation of post-RT ventilation loss. Specifically, we develop and validate a model, [P+A] to estimate post-RT ventilation based on radiation damage to lung parenchyma [P] and radiation-induced partial and complete obstruction of airways [A]. Materials/

Methods: Following informed consent, pre-RT breath-hold CTs (BHCTs) and pre- and post-RT 4DCTs were prospectively acquired from locally-advanced lung cancer patients receiving definitive treatment. For each patient, 9-12 generations of airways were segmented on the BHCT using a virtual bronchoscopy software. For each terminal airway, we defined a sub-lobar volume (SLV), comprising the nearest voxels within the containing lobe. 4DCT-based ventilation maps were calculated and used to estimate pre- and post-RT ventilation, serving as the model input and reference standard for model validation, respectively. For predictive modeling, ventilation loss due to parenchymal damage was estimated using a previous NTCP model. Ventilation loss due to partial or complete airway obstruction was estimated using our previous airway radiosensitivity model and fluid dynamics principles. We divided the data into training and test sets (Ntr and Nts in Table) and conducted a k-fold cross-validation by comparing predicted and actual post-RT ventilation in the SLVs in the test dataset and calculating receiver operating characteristic (ROC) metrics.

Results: Within a dataset of 560 SLVs from four evaluable patients, the P+A model outperformed the P model, with lower average RMSEs: 13 ± 3 percentile [P+A model] vs 24 ± 3 percentile [P model]. The P+A model aligned closely with the actual lobar post-RT ventilation maps derived from post-RT 4DCTs, particularly in high-dose regions. ROC analysis showed that the P+A model outperformed the P model by demonstrating higher sensitivity (0.98 vs. 0.07) and higher accuracy (0.87 vs. 0.25).

Conclusion: These results address an important gap in our understanding of the role of radiation-induced airway damage and its role in accurately modeling RILI.

Fold #	#Training SLVs (Ntr)	# Test SLVs (Nts)	RMSE P+A model (percentile)	RMSE P model (percentile)
1	452	108	15	23
2	411	149	14	28
3	398	162	10	22
4	419	141	14	20
		Average(+/-Std Dev)	13(+/-3)	24(+/-3)