Uncertainty estimation using a 3D probabilistic U-Net for segmentation with small radiotherapy clinical trial datasets

• DOI: 10.1016/j.compmedimag.2024.102403
• Publication Date: 2024-06-02T14:10:27Z
• Authors: Phillip Chlap; Hang Min; Jason Dowling; Matthew Field; Kirrily Cloak; Trevor Leong; Mark Lee; Julie Chu; Jennifer Tan; Phillip Tran; Tomas Kron; Mark Sidhom; Kirsty Wiltshire; Sarah Keats; Andrew Kneebone; Annette Haworth; Martin A. Ebert; Shalini K. Vinod; Lois Holloway
• Keywords: Ground truth; Uncertainty Quantification

Bio-medical image segmentation models typically attempt to predict one that resembles a ground-truth structure as closely possible. However, medical images are not perfect representations of anatomy, obtaining this ground truth is A surrogate commonly used have multiple expert observers define the same for dataset. When on there can be significant differences depending structure, quality/modality and region being defined. It often desirable estimate type aleatoric uncertainty in model help understand which true likely positioned. Furthermore, these datasets resource intensive so training such using limited data may required. With small dataset size, differing patient anatomy well represented causing epistemic should also estimated it determined cases effective or not.