Session Item

Organ-at-risk (OAR) delineation is a key step for radiotherapy treatment (RT) planning. Manual delineation of OARs is a tedious process, time consuming, and prone to errors due to intra- and inter-observer variations. The management of brain tumors typically involves RT planning based on computed tomography (CT) and magnetic resonance imaging (MRI). MRI is acquired for detailed tumor localization and delineations of the target and OARs thanks to its excellent soft-tissue contrast. In this study, we propose to use an automatic artificial intelligence-based OAR segmentation to support MR-based treatment planning for brain.

ART-Net®, a CE-marked, FDA-cleared   anatomically   preserving   deep-learning ensemble architecture for automatic contouring (AC) of OAR, was retrained on T1w-MRIs following the EPTN 2018 guidelines. For a total of 80 patients, automatic annotations of 25 OARs were performed and compared against the inter-expert variability between three expert annotators. In addition, 9 (chiasma, encephalon, left cochlea, left cornea, left eye lens, midbrain, posterior cerebellum, and right lacrimal gland) out of the 25 OARs were randomly selected and submitted to two independent observers for evaluation. Experts’ contours used for RT delivery were blended with the ones delineated by ART-Net® at a 50%-50% ratio. Random blending at the patient level was performed guaranteeing that, among contours being evaluated per patient and OAR, the 50%-50% split was satisfied. Experts were asked to score ACs and the contours from clinical experts (CE) as A/acceptable, B/ acceptable after minor corrections, and C/ not acceptable for clinical use. To avoid any bias, experts were blind to the origin of the contours (manual or AI).

Average Dice score coefficient (DSC) for ART-Net® on the testing set was 72.6% while that of the inter-expert variability was 69.1%. ART-Net® achieved mean dice greater than the inter-expert variability for all 25 organs (Tab.1). Encephalon reported the highest DSC 97.45 ± 0.50, while left vestibular and semi-circular canals (VSCC) reported the lowest DSC of 52.58 ± 14.55. These organs are normally not visible on T1Gd sequences, which might explain these results. 4 out of the 9 organs were evaluated as A for 100% of the cases. All 9 OARs achieved a 100% A+B score for the AI contours. A+B score for the manual delineation from experts was 100% for all but one organ (left eye lens).

In this study, we investigated the advantages of using a CE-marked, FDA-cleared anatomically   preserving   deep-learning ensemble architecture for automatic segmentation of OARs for brain RT on MRI. This work illustrates the strength of using ART-Net® for segmentation to support MR-based treatment planning in the brain as it consistently generates delineations with high clinical acceptability which can lead to reduction in variability of practice. Future work will include dosimetric evaluations.