Point-cloud segmentation with in-silico data augmentation for prostate cancer treatment

Background

In external x-ray radiation therapy, the administered dose distribution can deviate from the planned dose due to alterations in patient positioning, changes in intra-fraction anatomy, and the limited precision of the beam delivery system in spatial terms. Adaptive radiation therapy (ART) can potentially improve dose delivery accuracy by re-optimizing the treatment plan before each fraction, maximizing the dose to the target volume while minimizing exposure to surrounding radiosensitive organs. However, to effectively implement ART, the stages of the radiation therapy pipeline, including image acquisition, segmentation, physician directive generation, and treatment plan generation, must be optimized for maximum speed and accuracy to ensure feasibility prior to each treatment fraction. In this work, we focus on image segmentation. By reducing the segmentation computation time, one can reproduce the planning process for each session, enabling routine customization for individual patients, achieving safe dose escalation, better cancer control, and reduced risk of severe radiotoxicity.

Purpose

The aim of this study is to develop a fast point-cloud-based segmentation model with novel in-silico-aided data augmentation and demonstrate it on pelvic computed tomography (CT) patient data used in prostate cancer (PCa) treatment. This model can be implemented during ART because it requires only a few seconds to perform organ segmentation.

Methods

In this study, a dataset of pelvic CT images was obtained from Order of St. Francis (OSF) Healthcare Hospital (Peoria, IL, USA), comprising 38 images in total. These were divided into 25 for training, seven for validation, and six for testing the developed model. A novel point-cloud-based model was used to reduce the prostate segmentation time, cross-validation was implemented to ensure the robustness of the model. The developed point-cloud-based network is a novel deep-learning (DL) model that adds a loss function that combines region-based with a new boundary loss function. The region-based loss enables the identification of large volumes while the boundary loss, whose relative weight increases with the epochs, increases the network training ability of uneven surfaces, like the interface between the prostate bladder and rectum, which are challenging to resolve. We introduced a new data-augmentation approach to expand the training set. This fully automated method generates synthetic 3-D CT images by creating relevant organs in the extended cardiac-torso (XCAT) computational phantom. The Dice similarity coefficient was used as an assessment metric and compared to state-of-the-art segmentation models. The doses to the prostate and organs at risk (i.e., bladder and rectum) were also calculated for both our automated segmentation and manual expert segmentation to evaluate the practical feasibility of the point-cloud-based approach.

Results

Our model achieves the segmentation results (Dice coefficient) of 0.92 $\pm$ 0.04, 0.89 $\pm$ 0.05, and 0.84 $\pm$ 0.07 for bladder, prostate, and rectum, respectively. The accuracy of the prostate segmentation outperforms the voxel-based segmentation models reported in the literature. More importantly, the average segmentation time of the point-cloud model for a single 3-D CT data set was 1.8 times faster than 2-D fully convolutional network (FCN), and 11 times faster than 3-D U-Net. The improved loss function and in-silico-based training data augmentation approach effectively enabled the model to learn features of outlier data sets, thereby improving the model's robustness across diverse images. The developed fast and robust point-cloud segmentation model can potentially be applied to ART to improve the treatment workflow.

Conclusions

Our proposed method demonstrates favorable performance in the segmentation of X-ray CT data. Results confirmed that the point-cloud-model is faster than voxel-based segmentation algorithms while achieving comparable or better segmentation results. The segmentation approach can be integrated into ART workflow, ultimately reducing the workload of clinicians and radiologists.