Calibrated uncertainty estimation for interpretable proton computed tomography image correction using Bayesian deep...

Area of Science:

• Medical Imaging
• Computational Physics
• Radiotherapy Technology

Background:

• Proton computed tomography (pCT) is crucial for proton therapy planning, but suffers from noise and scatter, degrading image quality.
• Existing pCT correction methods often lack robust uncertainty estimation, which is vital for treatment accuracy.
• Accurate proton stopping power ratio (SPR) imaging is essential for precise dose delivery in proton therapy.

Purpose of the Study:

• To develop a novel uncertainty-aware pCT image correction method using a Bayesian convolutional neural network (BCNN).
• To enable accurate prediction of corrected SPR images and their associated uncertainties from noisy pCT data.
• To improve the reliability and safety of proton therapy by providing uncertainty quantification.

Main Methods:

• A DenseNet-based Bayesian convolutional neural network (BCNN) was developed to predict corrected SPR images and their uncertainties.
• Monte Carlo simulations generated 432 noisy pCT images from phantom data for training and testing.
• Heteroscedastic loss and deep ensemble techniques were employed to estimate aleatoric and epistemic uncertainties using 25 BCNN models.

Main Results:

• BCNN-corrected SPR images achieved high accuracy, reducing mean absolute error from 0.263 to 0.0538 compared to uncorrected images.
• The calibrated uncertainty estimates provided accurate confidence levels for the corrected images.
• BCNN-corrected water-equivalent thickness (WET) was significantly more accurate than that from a conventional non-Bayesian CNN method.

Conclusions:

• The proposed BCNN method effectively corrects noisy pCT images and quantifies image uncertainties.
• The estimated uncertainties can identify sources of SPR errors and inform the development of spot-specific range margins.
• This uncertainty-aware approach advances uncertainty-guided proton therapy, enhancing treatment precision and safety.