ARTIFICIAL INTELLIGENCE CONTOURING OF CARDIAC SUBSTRUCTURES IN RADIATION THERAPY: A COMPARISON OF SOLUTIONS

Abstract

Background and Purpose Artificial-intelligence (AI)-based contouring tools have emerged to improve the assessment of radiation doses to cardiac substructures beyond the mean heart dose. This study compares the raw results from multiple solutions and evaluates the impact of non-contrast enhancement on the contours for each solution.Materials and Methods Twenty lung cancer patients with both contrast-enhanced (CE-CT) and non-contrast enhanced (NCE-CT) thoracic scans, sequentially acquired during the same session, were analyzed. 7 commercial, 3 open-source, and one in-house AI solutions were compared on CE-CT using the Dice Similarity Coefficient (DSC) and the 95th percentile of Hausdorff distance (HD95). Volume ratios between NCE-CT and CE-CT were calculated to quantify the impact of non-contrast enhancement on contours for each solution.Results Typically, 10 cardiac substructures including the whole heart, are contoured by most of the solutions. For the whole heart, the cavities and the great vessels, median DSC is above 0.8 for approximately half of the structure-solution pairs, and median HD95 is below 1 cm for more than one third of these pairs. For the coronary arteries, median DSC is schematically either around 0.5 or close to 0 and median HD95 ranges between 1 cm and 7 cm.Non-contrast enhancement affected results variably across solutions and substructures. Except for one solution, volume differences are below 10% for most of the structure-solution pairs.Conclusions Automatic solutions provide inter-solution differences for cardiac substructures that could have clinical impact. Detailed information and standardization of the models, ideally through international consensus and shared datasets, are essential.

Note:
Funding Information: Dr Loïg Vaugier and Alexandra Moignier has a partnership with RaySearch Laboratories for detailed heart model development Dr Angela Botticella has a partnership with Therapanacea for detailed heart model development.

Declaration of Interests: None declared.

Keywords: artificial intelligence-based contouring, cardiac substructures, influence of contrast in CT images, radiotherapy, comparison of solutions