[Application of artificial intelligence-based organ contouring software and comparison of results in treatment planning for radiotherapy].

Abstract

Introduction: The first step in the radiotherapy treatment of cancer patients is the treatment planning, which involves delineations of the target volume (tumour) and organs at risk on a series of CT scans. This process is called contouring or segmentation. The delineation is performed in two dimensions on axial CT slices, but the dose distribution calculation and plan evaluation are carried out in three dimensions. Contouring has traditionally been done manually, but automatic organ contouring software are now available to make the process more accurate and consistent. Objective: To compare the quality of contouring software using artificial intelligence-based algorithms, and to define the organs that can be outlined with the best consistency, and those organs where there are large differences in segmentation between different software. Method: Contours of organs at risk defined by three contouring software (MVision, ART-Plan, Limbus) in four anatomical regions on a series of 93 CT scans were compared. MVision contours were the references, and contours defined by the two systems were compared with them based on volumes, centers of mass location, and spatial similarity and conformity indices. We ranked the spatial similarity, defining the organs with the smallest and largest differences. Results: Order of the organs with the best agreement: lungs, brain, liver, spleen, stomach, heart, eyes, mandible, kidneys, spinal cord, breasts, and bladder. The largest deviations were found at small-volume organs: thyroid, chiasma, left anterior descending artery, and pituitary. Large variations were found in the prostate and differences in the trachea were due to different interpretations of the anatomical boundaries. Discussion: For organs where the tissue density differs significantly from their surroundings (air–soft tissue, bone–soft tissue), the software can determine the organ in question with the appropriate quality. Small differences in contours were found for larger and relatively large differences for smaller organs. Conclusions: The contours were defined with good agreement by the software for most of the organs, large deviations were found only in small-volume organs. Orv Hetil. 2024; 165(49): 1934–1944.