A deep learning auto-segmentation tool for cardiac substructures in 4D radiotherapy planning for locally advanced lung cancer

The average scan from the 4D-CT dataset of 20 patients completing radical radiotherapy for lung cancer 2015-2020 at a tertiary centre were used for manual and automated cardiac substructure segmentation. All manual delineations were completed by a radiation oncologist and subsequently verified by a senior radiation oncologist and cardiologist. Scans were imported into MATLAB v2020b for auto-segmentation. Manual and automated substructures were geometrically compared by percentage volume difference (VD), centroid shift (CS), Dice similarity coefficient (DSC), and 95% percentile Hausdorff distance (HD95). The mean dose and maximum dose (D max) of the automated substructures were also compared against the corresponding manual dose. The two senior clinicians qualitatively assessed the performance of the auto-segmentation tool’s output.