Session Item

MRI based image-guided adaptive brachytherapy (IGABT) is an essential part of treatment of locally advanced cervical cancer (LACC), and considered state of the art. To monitor organ motion and its impact on delivered dose, routine control scans are taken before every second fraction. Applicator-based rigid registration of the two image series informs the decision if the current treatment plan needs to be adapted.

However, currently registering two image series on MRI is time-consuming. Semi-automatic routines exist but rely on prior definition of landmarks. The required time investment blocks the evaluation of inter and intra-fraction motions of organs at risk in clinical routine.

In this study we i) compare different applicator-based rigid image registration algorithms for MR-IGABT, and ii) train a neural network (NN) to predict the applicator structure to automate the registration process.

A cohort of 56 patients was available for this study. Patients were treated for LACC according to the EMBRACE2 protocol. An open MR-scanner (0.35T) was used to acquire the images, with in-plane resolution 256x256 (1.17 mm), and 5 mm slice thickness.

For each patient two BT fraction treatment plans, including MRI and reconstructed applicators were exported. The contours of the applicator, which are usually not available in the TPS (Oncentra Brachy, Elekta), were generated with an Elekta Applicator Slicer research plugin and treated as ground-truth masks.

Automatic registration of the image pairs was performed with five different registration algorithms written in Python. Tab. 1 provides an overview of the different registration configurations.

As ground-truth masks would not be available in clinical routine, a NN (UNET, implemented in PyTorch) was trained to predict the applicator structure in previously unseen MR-images. Training was performed using 5-fold cross validation and a random train-test split.

The best performing registration algorithm was re-run with the output of the NN, to compare the effect of ground-truth vs predicted masks.

Registration results were evaluated on the test set, using the root mean squared error of the applicator dwell positions (RMSE). Performance of the NN was measured using the DICE-coefficient.

The results of the different registration algorithms on the test set are summarized in Fig. 1. The best result is achieved by registering the distance maps generated from the ground truth applicator structures, resulting in an RMSE of 0.7 ± 0.5 mm. Using the predicted applicator structures an error of 2.7 ± 1.4 mm is achieved. Mean DICE of the predicted applicator masks is 0.7 ± 0.07.

This study shows that automatic applicator-based image registration for MR-IGABT could be achieved by combining classical image registration algorithms with modern deep learning methods. This is an important step towards future applications for monitoring organ motion during treatment and reducing dosimetric uncertainties.