ESTRO 2024 - Abstract Book

S3001

Physics - Autosegmentation

ESTRO 2024

Figure 1. Hausdorf distance (HD) for the left and right lungs compared to the gold standard for the ProSoma and Direct Organs contours and the HD for the heart contour for Direct Organs. The horizontal blue line indicates the target tolerance for an acceptable contour (2 mm, CT voxel size[2]). Data points overlaid in magenta.

Table 1. Mean DVH statistics for each organ for the gold standard, ProSoma and Direct Organs. Errors at the standard deviations, p values given in brackets.

Left Lung

Right lung

Heart

Contour

V12 Gy (%)

V12 Gy (%)

V2 Gy (%)

V10 Gy (%)

Mean dose (cGy)

Gold Standard

10.4 ± 2.8

11.0 ± 3.6

3.4 ± 2.6

0.2 ± 0.5

56.5 ± 22.3

9.1 ± 3.1 (0.08) 10.5 ± 2.9 (0.90)

9.7 ± 3.8 (0.12) 10.6 ± 3.5 (0.60)

ProSoma

3.6 ± 2.6 (0.62)

0.3 ± 0.5 (0.18)

Direct Organs

59.6 ± 23.6 (0.39)

Conclusion:

It has been shown that for the purpose of planning breast cancer treatments (tangential only) AI generated contours provide a sufficient estimate of the organ dose without correction. This has the potential to bring considerable workflow efficiency benefits to breast planning.

Keywords: deep learning, artificial intelligence, breast

References: