Robustness Analysis of a Deep Learning based sCT generation algorithm

Abstract

In the context of adaptive proton therapy, a treatment strategy that dynamically adjusts to anatomical changes throughout the course of radiotherapy, generating accurate synthetic CT (sCT) images from cone-beam CT (CBCT) scans remains a major challenge. This is primarily due to the presence of artefacts in CBCT and its limited accuracy in Hounsfield Units (HU), which compromises its reliability for dose calculations. Recent advances in deep learning have enabled promising approaches for direct CBCT-to-sCT translation. However, these methods typically lack robust mechanisms for quantifying uncertainty, which is essential for clinical decision-making, particularly in proton therapy, where dose distribution is highly sensitive to anatomical and HU variations. This thesis investigates the potential of two complementary state-of-the-art uncertainty quantification techniques to enhance the trustworthiness and interpretability of deep learning-based sCT generation models.

The study focuses on two main types of uncertainty: epistemic, which reflects model ignorance, and aleatoric, which captures data-inherent noise. Monte Carlo Dropout (MCD) and heteroscedastic regression are used to model each, respectively, within a U-Net architecture. The models are evaluated on a simulated dataset from IBA and on real patient data, which presented substantial artefacts and was preprocessed accordingly by IBA. Evaluation relies on multiple metrics, including MAE, RMSE, SSIM, PSNR, Expected Calibration Error (ECE), and the Pearson correlation between uncertainty and error (PCC).

In the simulated setting, MCD alone achieves an MAE of 21.08 ± 2.29 HU, and ECE of 14.63 HU, with a moderate correlation between uncertainty and error (PCC = 0.54). The combined model improves these metrics, reaching an MAE of 20.65 ± 2.00 HU, ECE of 4.41 HU, and PCC of 0.65. On the real dataset, where artefacts and noise are more prevalent, the combined model still improves over MCD alone though performance remains limited.

The discussion highlights that while MCD uncertainty maps aligns with general error structure in both datasets, they tend to be overconfident and insufficiently calibrated. The introduction of aleatoric modeling improves calibration and interpretability, particularly in identifying regions of anatomical ambiguity or noise. However, A failure to disentangle both uncertainty types still limit clinical applicability.

These findings suggest that while current uncertainty quantification methods such as MCD and heteroscedastic regression represent meaningful steps toward more reliable sCT generation, they remain insufficient to fully meet clinical trustworthiness and interpretability demands.