Session Item

Contour variation (inter- or intra-observer) and organ motion produce uncertainty in radiation dose delivery, which is challenging to quantify. Most automatic segmentation tools generate a single “most likely” contour, but other contours are possible, especially for small difficult-to-see structures. A technique is proposed to generate additional feasible auto-contours from visualisation and quantification of contour and motion uncertainties on lung cancer planning CTs, with evaluation of the dosimetric impact.

We previously developed a fully-automated segmentation tool to delineate 18 cardiac structures (Finnegan et al, Radiother Oncol 170(2022)S1,670): the heart via a deep learning (DL) model and substructures within the DL-heart using multi-atlas mapping (M-AM) and geometric segmentation. To map contour variation, the tool was used on 27 curative lung cancer planning CTs, with each of the 10 atlases from the M-AM step providing a separate substructure contour (Fig 1). For motion mapping, 9/27 cases had 4D planning CTs, and cardiac motion was mapped through all respiration phases. Dose volume histograms (DVH) for each probable contour and motion-changed contour were generated for comparison.

Dose differences from contour variation and motion depend on tumour position relative to organs at risk (OARs) and not simply on magnitude of contour differences. For example, motion mapping showed mean dose to right atrium (RA) and right coronary artery (RCA) differed by 0.7-7.4Gy and 0.5-5.9Gy respectively across the cases, with Fig 2A showing the case with high dose difference of these right sided structures (7.4Gy and 5.8Gy difference respectively). The right-sided tumour means right cardiac structures lie in a high dose gradient, hence wider dose differences in DVH. Conversely, left sided structures such as left ventricle (LV) and left anterior descending artery (LAD) for this case showed small difference (0.3-0.4Gy) in mean dose. The opposite is seen for left sided tumours as shown in Fig 2B’s example.

Doses vary considerably as a structure moves in and out of a high dose region, due to contour uncertainties and motion. In the clinical setting, an OAR’s vulnerability to such dose uncertainties can differ significantly between cases, as tumour and OAR positions relative to high dose gradient regions vary.  Automatic segmentation tools can be helpful to provide time-efficient and consistent contours, but being able to predict “uncertainty regions” adds significant further value. This work demonstrates the feasibility of predicting a “dose region” where the true dose most likely falls, rather than seeing DVHs as single curves per structure. This can flag vulnerable OARs – with a potential role in cardiotoxicity research on dose variations and outcomes, and to aid personalised plan optimisation in the future. While our automated tool is designed for cardiac structures, the methods described can also potentially be applied to other OARs.