ESTRO 2024 - Abstract Book

S3111

Physics - Autosegmentation

ESTRO 2024

The overlapping of the prediction and the groundtruth is promising in area and shape but the prediction has a smaller area than the groundtruth, and by observing other slices, this trend continues.

Conclusion:

Using a small amount of patients, we were able to achieve promising results for CTV lymph node auto-segmentation. Although the DSC values were high, the image shows that there is room for improving the quality of the predictions, this can be achieved by using pre-trained models on bigger medical datasets before fine-tuning for the lymph node segmentation. Further work should include clinical evaluation and verification if these methods are helpful in reducing the time that the RO spends doing the contouring of the images.

This study was part of the AuToMI project and funded by the Italian Ministry of Health GR-2019-12370739.

Table 1:

Fold

DSC

HD95 (mm)

1

0.82 ± 0.04

5.7 ± 1.53

2

0.83 ± 0.04

4.88 ± 1.95

3

0.84 ± 0.04

4.97 ± 1.6

4

0.82 ± 0.06

5.48 ± 1.87

5

0.83 ± 0.05

5.09 ± 1.39

Average

0.83 ± 0.05

5.23 ± 1.67