Logo of the Physikalisch-Technische Bundesanstalt

Using machine learning techniques for personalized CT dosimetry


Figure 1: Comparison of the ground truth (left) and the segmentation results of the U‑Net (right).

The aim of personalized computed tomography (CT) dosimetry is to obtain an immediate indication of the effective dose following CT imaging. The effective dose is regarded as the measure of the magnitude of radiation exposure with respect to stochastic radiation effects. As part of a doctoral project, a method is currently being developed to achieve personalized CT dosimetry based on techniques employed in the field of machine learning.

In order to quantitatively assess the stochastic radiation effects of CT, we need a measure that is based on both scanner‑specific and patient‑specific information. The currently used computed tomography dose index (CTDI) does not fully meet these criteria. In this context, the effective dose represents a better measurand, provided it is computed directly from the CT image. In contrast to the CTDI, the effective dose additionally allows a direct comparison between different X‑ray imaging modalities. Computing the effective dose from CT datasets is highly time- and labor‑intensive and as such not suitable for use in routine clinical practice. Our work is aimed at developing a procedure for rapid dose determination based on neural networks to achieve the computing speed required in the clinical workflow. In this procedure, the patient‑specific effective dose from a CT image is determined by solving two sub‑problems: multiorgan segmentation and dose simulation.

For multiorgan segmentation, we use a U‑Net convolutional neural network (CNN) architecture, which takes a two‑dimensional slice of a CT image and computes a segmentation for each of the organs contained in the training dataset. The training dataset currently consists of 140 CT scans with six segmented organs. One such segmentation is shown in Figure 1.

After training 50 epochs, multiorgan segmentation achieves an accuracy of 83 % on the validation dataset. The accuracy is given as the mean intersection over union between the segmentation performed by the network and the ground truth. Through an improved training dataset and hyperparameter tuning, we aim to achieve multiorgan segmentation with all the organs needed for calculating the effective dose.

The computation of the 3D dose distribution is to be done using a generative adversarial network. To generate the training data, a digital twin of PTB's research CT was developed based on Monte Carlo simulation. Initial experimental evaluations show deviations of less than 10 %. A simulated dose distribution is depicted in Figure 2. We expect that the methods described will permit the calculation of the effective dose within just a few minutes.

CT scan

Figure 2: Simulated dose distribution for a simulated CT scan for the ICRP phantom.


Marie-Luise Kuhlmann, Department 6.2, Working Group 6.25