ESTRO 2024 - Abstract Book

S3122

Physics - Autosegmentation

ESTRO 2024

USZ

166

35

0

Results:

The results of our study are summarized in Table 2. The integration of the spatial/anatomical information to the segmentation model reduced the total number of false positives and improved the precision for USZ. After applying the parcellation information, the total number of false positive metastases was reduced from 245 to 131. The precision also increased notably by 15%. Regarding F measures which consider the sensitivity and precision trade-off, integration of brain parcellation to the model could significantly improve the F measures for the USZ dataset, as a special spatial distribution near cortical surfaces and meninges was observed in this dataset. Considering the Erlangen and UCSF datasets, adding the brain metastases distribution did not have a significant impact on the results. This is potentially because of the coarse brain parcellation using linear registration. As a consequence, the model fails to learn the correlation between metastases spatial distribution and brain parcellation.

False positive rate

Total detected

Data-Method Total BMs

Total FP

Sensitivity Precision

F1 score

F2 score

Erlangen without percolation

272

234

27

0.859

0.899

0.25

0.878

0.866

Erlangen-with parcellation

272

232

24

0.853

0.906

0.23

0.879

0.863

UCSF-without parcellation

3340

3099

718

0.928

0.812

2.23

0.866

0.902

UCSF-with parcellation

3340

3112

758

0.932

0.804

2.34

0.863

0.903

USZ-without parcellation

114

124

245

0.861

0.336

7.0

0.483

0.656

USZ-with parcellation

114

125

131

0.868

0.488

3.74

0.625

0.751

Conclusion:

Based on the results of our study, integrating the brain metastases location information can be beneficial in the detection performance of deep learning models and in reducing false positive numbers for datasets with special spatial metastase distribution. Otherwise, no significant benefit is observed.

Keywords: Brain metastases, deep learning