ESTRO 2024 - Abstract Book

S3007

Physics - Autosegmentation

ESTRO 2024

reflecting less consistency. On qualitative evaluation, the b-AI and manual delineations needed few or no corrections in 69-100% of delineations, whereas the c-AI achieved this level of accuracy in only 27-63% of delineated OARs. Quantitative and qualitative evaluations overall agreed, apart from spinal cord, where some manual delineators produced very large spinal cords, in conflict with the consensus definition used to generate the b-AI. Doses to OAR using b-AI contours or STAPLE contours are shown for selected metrics in figure 2. For all metrics, high agreement was seen with correlation coefficients ranging from 0.95 to 1.00. However, some outliers were observed, especially in D1cm3, reflecting the sharp dose gradients in the dose-escalated plans.

OAR

Manual

b-AI

c-AI

Manual

b-AI

c-AI

Median SDSC [IQR] 0.88 [0.84 0.93] 0.94 [0.89 0.97] 0.94 [0.87 0.98] 0.88 [0.82 0.91] 0.91 [0.85 0.92] 0.97 [0.96 0.98] 0.97 [0.94 0.98] 0.92 [0.83 0.94] 0.80 [0.63 0.89]

median SDSC [IQR] 0.89 [0.88 0.90] 0.96 [0.93 0.96] 0.90 [0.86 0.96] 0.83 [0.77 0.84] 0.88 [0.82 0.90] 0.93 [0.90 0.94] 0.94 [0.90 0.96] 0.88 [0.85 0.93] 0.63 [0.48 0.72]

median SDSC [IQR] 0.74 [0.67 0.83]

% No/few corrections

% No/few corrections

% No/few corrections

Aorta

85.2

97.6

63.6

Bronchi

N/A

73.8

76.2

N/A

Trachea

N/A

88.8

100.0

N/A

0.76 [0.73 0.79] 0.70 [0.65 0.72] 0.90 [0.87 0.92] 0.91 [0.88 0.93]

Esophagus

89.3

93.0

61.8

Heart

79.5

88.1

27.3

Left Lung

81.0

69.0

33.3

Right Lung

89.0

69.0

45.5

Spinal Canal Spinal Cord

N/A

70.5

97.6

N/A

0.81 [0.79 0.87]

67.8

97.6

57.6

Table 1: Median 1 mm surface dice similarity coefficients (SDSCs) for manual, b-AI and c-AI contours compared to STAPLE ground truth delineations, and proportions of delineations needing only no or few corrections depending on contouring method. IQR: Inter Quartile Range. N/A: Not available.

Figure 1: Delineations of thoracic organs at risk generated by the b-AI