ESTRO 2024 - Abstract Book

S3036

Physics - Autosegmentation

ESTRO 2024

Figure 1. Examples of uncertainty maps. In each subfigure A and B, the first row shows the PET/CT images (axial, sagittal and coronal directions) while the second row shows the CT image with the associated uncertainty map overlaid. The ground truth (cyan) and predicted (blue) contours are also shown.

Figure 2 . Box plot of the entropy of True Positive, False Positive and False Negative voxels. Entropy values are within the range [0, 0.37]. Voxels with higher entropy indicate higher uncertainty of their predicted class (GTV or not GTV). The red dots indicate the mean values. The rectangular boxes show the interquartile range (25th, 50th and 75th percentile), while the lower and upper whiskers indicate the 5th and 95th percentiles, respectively. Outliers are represented as black dots outside the 5th and 95th percentiles.

Conclusion:

In conclusion, uncertainty estimation using Monte-Carlo dropout can potentially be used as a quality control step for deep learning-based auto-segmentation tasks. Highlighting regions where the automatic contours are uncertain may enable the clinician to focus on these areas during contour revision, potentially improving the overall contouring efficiency further.

Keywords: head and neck, deep learning, uncertainty

References:

Damianou, A., & Lawrence, N. D. (2013). Deep gaussian processes. In Artificial intelligence and statistics (pp. 207-215). PMLR.

Gal, Y., & Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning (pp. 1050-1059). PMLR.

Milletari, F., Navab, N., & Ahmadi, S.-A. (2016). V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In 2016 Fourth International Conference on 3D Vision (3DV). 2016 Fourth International Conference on 3D Vision (3DV). IEEE. https://doi.org/10.1109/3dv.2016.79