ESTRO 2024 - Abstract Book

S3039

Physics - Autosegmentation

ESTRO 2024

The median DSC ranged from 0.85 to 0.97. The lowest median DSC was observed for the right optic tract[YAL1] , while the highest median DSC was found for the brainstem. As for the medians of NDSC1mm, they fell within the range of (0.90-0.96), with the lowest value belonging to the right hippocampus and the highest median, to the left optic tract. The HD95 served as the distance metric, and across all organs, the medians ranged from 1.0 to 2.3. The minimum value recorded was 1.0 for the right optic tract, and the maximum value reached 16 for the left optic tract. Results of all evaluated predictions are presented in Figure 1.

Variance

NSD1mm showed a moderately consistent median of variances ranging from 0.0011 to 0.0324 across different organs, indicating the algorithms' stability. Slightly higher variances of the DSC coefficient were acquired with medians spanning from 0.0001 to 2.2 across all organs. Variance measures for HD95 were computed in a similar manner across five folds, resulting in median variances of 0.041 to 38.61. The results from all evaluated predictions are presented in Figure 1.

Conclusion:

In this study, we evaluated a semantic segmentation algorithm that combines CNNs and Transformers for brain OARs, in terms of model quality and consistency.

Results are indicative of robust and consistent delineation of brain OARs when combining the predictions of the transformer and CNN model, illustrated by high median DSC, NSDSC1mm and low HD95s and variances for predictions across the five folds of brain OARs.

Keywords: Transformers, convolutional network

References:

[1] Turcas, Andrada, et al. "Deep-learning magnetic resonance imaging-based automatic segmentation for organs at-risk in the brain: Accuracy and impact on dose distribution." Physics and Imaging in Radiation Oncology (2023): 100454. [2] Taghanaki, Saeid Asgari, et al. "Combo loss: Handling input and output imbalance in multi-organ segmentation." Computerized Medical Imaging and Graphics 75 (2019): 24-33. [3] Smith, Abraham George, et al. "Localise to segment: crop to improve organ at risk segmentation accuracy." arXiv preprint arXiv:2304.04606 (2023).] [4] Torpmann-Hagen, Birk, et al. "Segmentation Consistency Training: Out-of-Distribution Generalization for Medical Image Segmentation." 2022 IEEE International Symposium on Multimedia (ISM). IEEE, 2022.

1171

Digital Poster

Automated paediatric vertebrae landmark detection & growth measurement: AI vs analytical approaches

Kartik Kumar 1,2 , Adam Yeo 1,2,3 , Tomas Kron 2,1,3 , Rick Franich 1,2