ESTRO 2024 - Abstract Book

S4546

Physics - Machine learning models and clinical applications

ESTRO 2024

Data-dependent (aleatoric): incompleteness in input data (noisy inputs, imprecise labels) Model-dependent (epistemic): uncertainty associated with variability in model parameters.

The purpose of this study was to introduce a CBCT-to-CT generation method that provide sCT and predict its associated uncertainties maps by comparing them with the absolute error map (sCT/CT). This approach was aimed at incorporating sCT uncertainties into the daily dose calculation process in the context of the ART.

Material/Methods:

For this study, 85 patients with locally advanced oropharyngeal carcinomas treated with external beam RT were retrospectively selected. A planning CT was acquired for each patient, and for 14 patients a weekly CT was acquired during treatment course. All CT were acquired on a BigBore scanner (Philips), with a slice thickness of 2 mm. CBCTs were acquired with an XVI (Elekta) on a VersaHD linac (Elekta).

For each patient, the CT and CBCT were cropped to the smaller CBCT field of view and resized to 256*256*128 voxels using B-Spline interpolation.

sCTs and uncertainty heatmaps were generated by a 3D conditional GAN [2] composed of two adversarial networks (the generator and the discriminator). The generator (G) was a custom 6-block residual network, situated between a down-sampling block and an up-sampling block, featuring two distinct branches, each consisting of two convolutional layers to predict outputs. The discriminator was a 70³ PatchGAN.

To quantify data uncertainties, we assume Laplace-distributed image noise [3] and task the network with minimizing the following loss to estimate its variance σ(x): L unc = [|CT-G(CBCT)|/σ(x)]+log(σ(x))

σ(x) represents the data-based uncertainty map.

To quantify model-based uncertainties, the Monte Carlo dropout method was used, approximating Bayesian inference and characterizing predictions as a Gaussian process.

Once the estimation of uncertainties, a calibration process was employed to correct dropout-based uncertainties that lack proper calibration. This involves aligning them with the real errors observed in our training dataset through linear regression. To prevent an overestimation of error, both uncertainties were combined through quadratic summation, as the occurrence of two extreme error values concurrently is highly improbable. Total uncertainty maps (model-based +data-based uncertainties) were compared with the absolute error map using the Pearson correlation coefficient (PCC). To assess the robustness of the uncertainty estimation method, three patients from different centers than the training center [4] were included in the evaluation.

Results:

Figure 1 shows four significant results: