Table 1.
Patient characteristics in this trial cohort.
Fig 1.
Representative example of an automatic segmentation (blue) compared to manual contouring (red) of the whole heart.
The imaging volume shown corresponds to the automatically cropped region, defined using the bounding box of the lungs.
Table 2.
Quantitative comparison of the automatic whole heart segmentation to manual contouring.
Fig 2.
Planning organ at risk volumes (PRVs) for the heart, generated by manually contouring the maximum intensity projection image manually (left) and the union of automatic segmentations on each breathing phase (right), shown here as surface meshes in sagittal projection.
Our automatic approach typically generates smoother and more anatomically consistent delineations.
Fig 3.
A representative example of the differences in automatic segmentation using our proposed method (blue contour), propagation of delineations from the 0% phase using deformable image registration, and independent automatic segmentation for each individual phase (orange contour), compared to manual contouring (blue contour).
This figure illustrates results for the 40% phase image, where maximum deviation would be expected. There are minor variations at the base and apex of the heart.
Fig 4.
A comparison on automatic segmentation using our proposed method, propagation of delineations from the 0% phase using deformable image registration, compared to independent automatic segmentation for each individual phase.
We use the Dice similarity coefficient (DSC), mean absolute surface-to-surface distance (MASD) and Hausdorff distance (HD) to compare the automatic segmentations derived using either method to the manual contouring. Results are presented as box plots aggregated over the patient cohort. Overall, both methods provide similar performance, with no statistically significant difference in delineation accuracy.
Fig 5.
Examples of results from our proposed automatic heart segmentation method (blue), compared to manual contouring (red), shown here for the best (top), median (middle) and worst (bottom) performing patient image sets, as ranked using the mean absolute surface-to-surface distance (MASD).
We display each binned respiratory phase image, 0% - 90% in 10% increments going left to right. Overall, our method provides consistently accurate delineations. Some variations from manual contouring can be seen, for example in the 60% in the bottom row, where deviations in manual contouring can be seen on the borders of the heart.
Fig 6.
A demonstration of an evaluation of the dosimetric impact of using the automatic heart segmentation (Heart_auto, brown) relative to the clinical heart contour (Heart_manual, pink).
We observe a higher overall dose to the automatically defined volume, which was generated as the union of delineations on each individual binned respiratory phase from the 4D-CT image. Top: contours for the heart and tumour (red) shown in an axial slice on the maximum intensity CT image, bottom: a dose-volume histogram for the planning target volume (PTV) and several organs at risk.