Main Session
Sep 29
PQA 07 - Head and Neck Cancer, Lung Cancer/Thoracic Malignancies, and Nursing and Supportive Care

3588 - AI-Based Contouring and Dosimetric Validation of Cardiac Substructures in Stage II-III NSCLC in a Statewide Consortium

03:45pm - 05:00pm ET
Poster Hall - Exhibit Hall A
Screen: 25
POSTER

Presenter(s)

Travis Pflederer, MD Headshot
Travis Pflederer, MD - University of Michigan Radiation Oncology, Ann Arbor, MI

T. Pflederer1,2, D. Polan2, J. I. Matchynski1,2, A. Dubuque2, C. Hadley2, R. Jiang3, H. Yin2,3, N. Piersma2, M. Grubb2, R. Marsh2, M. M. Matuszak2, M. Schipper2,3, and S. Jolly2; 1Department of Radiation Oncology, Veterans Affairs Ann Arbor Health System, Ann Arbor, MI, 2Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, 3Department of Biostatistics, University of Michigan, Ann Arbor, MI

Purpose/Objective(s):

Whole heart dose metrics predict cardiac events and survival in locally advanced NSCLC. More recently, links have been made between coronary/chamber dose to major adverse cardiac events and all-cause mortality. Leveraging a statewide consortium with prospectively collected DICOM data, we utilized a previously trained and validated deep learning-based cardiac substructure segmentation model to (1) validate AI-generated contours against physician contours in this new multi-institutional dataset and (2) report cardiac substructure dose metrics in an initial sub-cohort.

Materials/Methods:

Eligible patients had stage II-III NSCLC treated with concurrent chemoRT (54-70Gy, 1.8-2.0 Gy/fx) from 2018-2023 with a planning target volume within 2 cm of the heart and =1 follow-up cardiac hospitalization forms filled out at each follow-up visit and annually. 872 patients met eligibility criteria. From 130 randomly selected patients, 43 underwent manual cardiac substructure contouring by a radiation oncologist. AI-generated contours were compared using DICE similarity coefficient (DSC), average symmetric surface distance (ASSD), and true Hausdorff distance (HD). Cardiac substructures contoured included the left atrium (LA), left ventricle (LV), right atrium (RA), right ventricle (RV), left main coronary artery (LM), left anterior descending artery (LAD), left circumflex artery (LCX), right coronary artery (RCA), and posterior descending artery (PDA). An additional 43 AI-contoured cases underwent focused manual review for gross segmentation errors, defined as > 1 cm.

Results:

Among 43 manually contoured cases, median whole heart DSC/ASSD/HD were 0.94/1.7 mm/14.2 mm. The values for the substructures are reported in the table below. Median absolute differences between AI- and manual contours were 5.0% for LAD V15Gy and 0.6 Gy for mean heart dose. Of the 43 AI-contoured cases reviewed, 2 (4.7%) showed gross errors: one missing part of the proximal RCA (2-3 cm) and one missing part of the proximal LAD (3-4 cm). In the 130-patient sub-cohort, median mean heart dose was 9.0 Gy; median mean chamber doses (LA/LV/RA/RV) were (13.5/2.7/7.3/3.1 Gy); the median V15Gy coronary doses (LAD/LCX) were (1.9 %/0.9%).

Conclusion:

AI-based cardiac substructure contouring demonstrated high geometric agreement for the chambers and whole heart, with expected lower geometric concordance for small coronary substructures in a statewide NSCLC cohort. Findings support using this scalable approach across the statewide cohort to enable rigorous evaluation of cardiac substructure dose-outcome relationships using prospectively collected hospitalization data.
Name

 LA

 LV

 RA

 RV

 LM

 LAD

 LCX

 RCA

 PDA
DSC

 0.81

 0.89

 0.83

 0.76

 0.22

 0.45

 0.29

 0.35

 0.15
ASSD (mm)

 2.4

 2.0

 2.1

 3.6

 3.5

 2.1

 3.1

 3.3

 5.1
HD (mm)

 14.0

 13.9

 15.6

 23.3

 10.4

 12.7

 13.8

 16.0

 19.0