Click here to

Session: Quantitative Imaging [Return to Session]

Automating 4DCT-Ventilation Imaging Generation Using AI-Based Advanced Lung Contours

Y Chen1*, S Pahlavian2, P Jacobs2, F Forghani3, E Castillo4, R Castillo5, Y Vinogradskiy1, (1) Thomas Jefferson University, Philadelphia, PA, (2) MIM Software Inc., Beachwood, OH, (3) Washington University, Creve Coeur, MO, (4) University of Texas at Austin, Austin, TX, (5) Emory University, Atlanta, GA


WE-C1000-IePD-F1-3 (Wednesday, 7/13/2022) 10:00 AM - 10:30 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 1

Purpose: Functional avoidance radiotherapy uses functional imaging to reduce pulmonary toxicity by designing radiotherapy plans that reduce doses to functional lung. A novel form of lung functional imaging applied for functional avoidance uses 4DCT imaging to calculating 4DCT-based lung ventilation (4DCT-ventilation). The process of generating 4DCT-ventilation images requires advanced lung segmentation (consisting of a standard lung with airway and vasculature removed) which is a manual and time-consuming process. The purpose of this work is to automate 4DCT-ventilation imaging generation using AI-based auto-segmentation techniques for advanced lung contouring.

Methods: 429 patients with 4DCT data from two institutions were used. Three methods for lung contours were generated including two advanced (with airway and vasculature removed): 1) manual segmentation (‘Lung-Manual’), 2) AI-based contours (‘Lung-AI’) and 3) AI-based standard lung contours used for treatment planning (‘Lung-RadOnc’). The AI model based on a residual 3D U-Net was trained using Lung-Manual of 356 patients. The predicted Lung-AI were validated against Lung-Manual contours using 73 independent patients with Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95). 4DCT-ventilaton images were calculated using all three contours and the images generated with the Lung-RadOnc and Lung-AI contours were compared against images generated with Lung-Manual contours (Pearson correlation).

Results: The DSC and HD95 comparing Lung-AI and Lung-Manual contours were 0.95±0.02 and 3.61±15.97 mm, respectively. The correlation between 4DCT-ventilation images generated with Lung-AI and Lung-Manual contours was 0.83±0.17, while the correlation between Lung-RadOnc and Lung-Manual-based imaging was 0.48±0.14.

Conclusion: Our study shows that using standard lung contours can result in inaccurate 4DCT-ventlation images and using AI-based advanced lung contours can produce 4DCT-ventilation images highly-correlated to those generated using manual (and time consuming) methods. The presented study uses a large patient database to automate the 4DCT-ventilation image generation process, which facilitates the integration of this novel imaging modality into busy clinical environments.

Funding Support, Disclosures, and Conflict of Interest: Funded by NCI RO1CA236857 and research agreement with MIM software


Lung, Ventilation/perfusion, Functional Imaging


IM/TH- Image Analysis (Single Modality or Multi-Modality): Quantitative imaging

Contact Email