ESTRO 2024 - Abstract Book

S3113

Physics - Autosegmentation

ESTRO 2024

2588

Digital Poster

Complete Uncertainty Estimation and Quantification In Automatic Prostate Cancer OAR Segmentation

Sergio Carreras-Salinas 1 , Lucía Cubero 1,2 , Carlos Sosa-Marrero 2 , Renaud De Crevoisier 2 , Oscar Acosta 2 , Javier Pascau 1,3

1 Universidad Carlos III de Madrid, Departamento de Bioingeniería, Madrid, Spain. 2 Université Rennes, CLCC Eugène Marquis, Rennes, France. 3 Instituto de Investigación Sanitaria Gregorio Marañón, Aplicaciones biomédicas de la ingeniería, Madrid, Spain

Purpose/Objective:

In recent years, Deep Learning has been adopted as a tool for prostate cancer organs at risk (OAR) segmentation. However, Deep Learning models exhibit a pronounced susceptibility to artifacts, data noise, variations in anatomical boundaries, image quality and contrast, and out-of-distribution patient data. These factors collectively contribute to the emergence of ambiguities within the segmentation models’ predictions, which are formally defined as uncertainties. The uncertainty inherent to the model’s lack of knowledge is defined as epistemic uncertainty, whereas the uncertainty within the data itself is called aleatoric uncertainty. Understanding and quantifying these uncertainties is crucial to determine how robust and trustworthy an OAR segmentation model is. In this context, this study aimed to estimate and quantify epistemic and aleatoric uncertainties for Deep Learning-based prostate cancer OAR segmentation.

Material/Methods:

The database included 172 prostate cancer patients with acquired 3D CT images and the respective manual segmentations for the prostate, bladder, rectum and seminal vesicles. Two different strategies were used to estimate epistemic and aleatoric uncertainties independently. Monte Carlo Dropout (MCDO) with a 3D U-Net as backbone was implemented to estimate the epistemic uncertainty. On the other hand, the estimation of the aleatoric uncertainty was conducted by training a Probabilistic U-Net (PU-Net), which combines a 3D U-Net with a variational autoencoder to introduce stochasticity in the predicted segmentations. First, five different MCDO models were trained on separate data folds. Since the PU-Net requires multiple segmentations for each patient to capture the aleatoric uncertainty, the trained MCDO models were used to generate four additional pseudo-labels for each OAR, to assemble an augmented database with five different pseudo ground-truth contours per OAR for each patient. These data were finally used to train the PU-Net. All MCDO models were trained for 120 epochs with a learning rate of 0.001 and the Generalized Dice Loss combined with Cross Entropy as loss function. The PU-Net was trained for 200 epochs, an exponential learning rate of 0.001 with gamma of 0.96 and Evidence lower bound (ELBO) as loss function using the Generalized Dice Loss as reconstruction loss and a beta equal to 10. For both MCDO and PU-Net models, AdamW was used as optimizer and they were trained on a Nvidia RTX 3090 GPU with 24GB of memory. During inference, for each CT image and OAR, 50 different segmentations were generated by both MCDO and PU-Net models, allowing to estimate the epistemic and aleatoric uncertainties, respectively. To quantify these uncertainties, the standard deviation, averaged ensemble Kullback-Leibler (KL) divergence and predictive entropy between the 50 segmentations were calculated.

Results: