ESTRO 2024 - Abstract Book

S2997

Physics - Autosegmentation

ESTRO 2024

Region

Model

Box

Points

Box+points

HN

ViT-H

73.3%

85.3%

88.7%

HN

LVM-Med

73.5%

86.2%

90.0%

HN

MedSAM

53.3%

53.3%

47.4%

Pelvis

ViT-B

89.1%

92.7%

96.1%

Pelvis

ViT-H

89.4%

93.1%

96.1%

Pelvis

LVM-Med

87.7%

93.1%

95.8%

Pelvis

MedSAM

57.3%

67.6%

65.2%

Table 1: Evaluation of prompting strategies (mean DSC).

Box prompt was the least favorable prompting strategy (Table 1), however, it was equivalent to 5.07 number of clicks when using points with ViT-B. Box+points prompt provided significantly higher scores compared to other strategies in almost all cases (Fig. 1). The only exception is the MedSAM model, where the mask decoder was trained using only box. If a box prompt (defined by 2 clicks) was followed by 8 more points, the DSC was still higher by 3-4% compared to 10 points prompting, depending on the model. In general, ViT-B and LVM-Med performed the best head-to-head. Their maximum DSC difference across all strategies was 1.39%, but usually marginal. ViT-B achieved DSC 90.06% for HN and 96.13% for pelvis region that was 9.5% and 16.57% higher compared to the performance of the SegResNet. ViT-H despite being computationally more expensive, provides inferior values.

Finally, we checked whether the refinement step increases the accuracy. For box prompt, refinement always increased DSC by at least 1.51%. For other strategies the accuracy was not affected significantly and rarely increased.