Session Item

One of the key aspects of the daily on-line adaptive radiotherapy (DART) workflow is to ensure fast and accurate contours on the daily imaging. Convolutional neural networks (CNNs) were used to model a dense displacement field (DDF) which propagate the planning structures to the new daily anatomy. Although these methods are fast, none of them produce interpretable uncertainties of the DDF, leading to a careful review of the contours by the physicians. The purpose of this work is to develop a method which can quickly generate the daily structures along with its interpretable uncertainties to focus physician attention to critical regions and thus reduce intervention time.

In this work, CNNs combined with drop-out layers were used to generate a distribution of the DDF which maps the planning CT to the daily CBCTs in head and neck treatments. Daily probability maps of the position of all the OARs and targets were generated averaging the planning structures propagated with 100 samples of the DDF distribution. The performance of the method was assessed by calculating the dice similarity coefficient (DICE) and the target registration error (TRE) between structures propagated with the mean DDF and their daily ground-truth. In addition, the results were compared to those obtained using a different DIR method (Elastix). The predicted DDF uncertainty was assessed as the isolation of two source of uncertainties; image uncertainty (related to regions of the image with very low contrast tissue), and model uncertainty (related to an improper matching of the two images).

It was found that the DICE and TRE distributions calculated with both methods (Elastix and the presented method) were not statistically significantly different for all the 11 structures evaluated in each patient of the test set. The esophagus structure had the lowest mean dice coefficient (0.73) and the maximum mean TRE (3.7 mm). Our method performed the registration in the order of seconds while Elastix required minutes. The mean standard deviation of the DDF obtained in the registration of the test images was 0.43 mm. The uncertainty analysis showed that our method correctly estimated the model uncertainty while underestimated the image uncertainty.

In summary, a fast and accurate DIR method which predict the DDF uncertainty was developed using CNNs and drop-out layers. It was shown that the accuracy of the method to generate the daily structures was not statistically different to Elastix. The uncertainty analysis has shown that this network was able to estimate the model uncertainty but underestimates the image uncertainty. Future works must improve DDF uncertainty prediction and evaluation to facilitate physician’s review of the structures and save time in a daily on-line adaptive workflow.