Click here to


Are you sure ?

Yes, do it No, cancel

Uncertainty-Aware Reconstructed Image Correction for Proton Computed Tomography Using Bayesian Deep Learning

Y Nomura1*, S Tanaka1, J Wang1, H Shirato1, S Shimizu1, L Xing1,2, (1) Hokkaido University, Sapporo, Hokkaido, Japan, (2) Stanford University, Palo Alto, CA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Integrated-type proton computed tomography (pCT) measures proton stopping power ratio (SPR) images for proton therapy treatment planning, but its image quality is degraded due to noise and scatter. Although several correction methods have been proposed, the correction methods that also estimate uncertainty are limited. This study aims to propose a novel image correction method with uncertainty quantification using a Bayesian convolutional neural network (CNN).

Methods: A DenseNet-based CNN was constructed to predict both noise-free SPR image and its uncertainty by using a noisy SPR image as input. 288 noisy SPR images of 6 non-anthropomorphic phantoms were collected with Monte Carlo simulations, while true images were calculated manually using known geometry and chemical components. Bayesian ensemble technique was performed to estimate aleatoric and epistemic uncertainties by training 32 CNN models with different initial random seeds. 200-epoch end-to-end training was implemented for each model with data augmentations of random volume flip and 90° rotations. For evaluation of trained model, accuracy of CNN-corrected head phantom images was compared with the actual images.

Results: The CNN-corrected SPR images represented noise-free images accurately. Mean absolute error in head phantom images was improved from 0.209 to 0.0600. Moreover, the calculated aleatoric and epistemic uncertainties were well correlated with an absolute difference between output and label. Computation time for calculating one image and its uncertainties with the ensemble of 32 CNN models is around 2.37 seconds with a NVIDIA RTX 2080Ti GPU.

Conclusion: A novel pCT image correction method was established using a Bayesian CNN. Our model is able to predict accurate pCT images as well as two types of uncertainty in nearly real-time. These uncertainties will be useful to identify potential cause of range errors and develop a patient-specific or spot-specific range margin criterion. This technique will be tremendously valuable in image-guided proton therapy.


Protons, CT, Bayesian Statistics


IM- Particle (e.g., Proton) CT: Machine learning, computer vision

Contact Email