Session Item

Accurate and consistent delineation of cardiac substructures is challenging. Manual delineation is labour intensive, with intra- and inter-observer variations, particularly where structures are difficult to visualise on CT. Automatic contours can be of limited accuracy when anatomical variations are present, as is often the case in lung cancer images. These are major barriers to cardiac substructure dose calculation.  Our aim is to validate a robust automatic segmentation tool that can accurately calculate cardiac substructure doses, with potential application in clinical settings and in large cohort studies.

A novel 18 cardiac structure hybrid automatic segmentation method was developed, combining deep learning segmentation of the whole heart, and mapping of 17 cardiac substructures using deformable registration from a multi-atlas set (see ESTRO 2022 Physics Track abstract, Finnegan et al). This was validated on 30 lung cancer cases which included anatomical and imaging variations, such as tumour mass abutting heart, lung collapse, metal and motion artefacts. Comparisons between manual contours of the 18 structures and automatic segmentations were performed using Dice similarity coefficient (DSC), and average difference in surface distance (mean distance to agreement, MDA). Radiotherapy dose difference was evaluated in 27 cases with planned dose ≥50Gy (range 54-66Gy). A novel tool was developed to calculate predicted dose ranges for cardiac substructures, by modelling volume and dose uncertainties.

Comparison of manual and automatic contours across all cases showed median DSC of 0.75-0.93 and median MDA of 2.09-3.34mm for the whole heart and four chambers. Median MDA for great vessels, coronary arteries, cardiac valves, sinoatrial and atrioventricular conduction nodes were 3.01-8.54mm.

Cases with planned dose ≥50Gy showed good dose correlation between manual and automatic contours (Fig.1). Median dose difference (manual contour dose minus automatic contour dose) was -1.23Gy to 0.51Gy for mean dose to heart and chambers; and -2.27Gy to 4.94Gy for mean dose to substructures. For maximum dose received, the median dose difference was -3.51Gy to 0Gy for heart and chambers, and -4.35Gy to 1.03Gy for smaller substructures.

Modelling volume and dose uncertainties of the automatic segmentation method enabled  predicted dose regions to be applied to dose-volume histograms, which can aid clinicians in considering the likely dose range that cardiac substructures will receive for a given radiotherapy plan (example shown Fig.2).

This novel hybrid deep learning tool has shown high accuracy and consistency in a validation set of lung cancer cases with challenging anatomical and imaging variations. By modelling dose uncertainty, this has promising applications in substructure dose calculations, radiotherapy plan optimisation, and feasibility of large cohort studies. Standardisation of use could also prove valuable for future studies on long-term cardiac toxicity.