PP08 Presentation Time: 10:40 AM: Deep-Learning Based Contouring of TRUS Images with Implanted Needles for HDR Prostate Brachytherapy

Abstract

<h3>Purpose</h3> In transrectal ultrasound (TRUS)-based high-dose rate (HDR), the radiation treatment plan is created after needle insertion. However, prostate contouring can be challenging, due to poor soft-tissue contrast, bleeding, and imaging artifacts from implanted needles. In recent years, deep learning (DL) algorithms have shown considerable promise in providing prostate contours from TRUS images, mostly without implanted needles. We hypothesize that a DL algorithm can provide high quality prostate contours from TRUS images with implanted needles. <h3>Materials and Methods</h3> We conducted a single-institutional retrospective study of 150 patients, who underwent TRUS-based HDR BT between 2020 and 2021. The dataset was randomly split into training (n = 110), validation (20), and testing (20) datasets. A 3D UNet with a resNet encoder was trained to provide automated prostate contouring. Since the clinical target volume for a subset of cases also included proximal seminal vesicles, all automated contours were post-processed to ensure that final contours only included the prostate. For the testing dataset, we evaluated the Dice coefficient between the DL and reference contours. Furthermore, for a 10-patient subset, we evaluated whether there were any significant differences in Dice coefficients between the reference and 5, independently acquired, contour sets (1 DL, 4 contour sets from radiation oncologists). <h3>Results</h3> For the testing dataset, the median dice coefficient was 0.87 (IQR=0.04). Automated contouring took (0.98±0.53) seconds. For the 10 cases subset the median dice for DL based contouring was 0.88 (IQR=0.05). Among all radiation oncologists manual contouring led to a dice of 0.83 (IQR=0.12). A boxplot of the contouring performance with respect to dice can be found in Fig. 1a. Contours for a central axial slice for a representative patient can be found Fig. 1b. The Pearson's correlation between the DL generated and reference prostate volumes was r=0.98. For manual contouring the correlation was r=0.93 (IQR=0.06). Among the 10-patient subset, there was no significant difference (p>0.5) in Dice coefficients with respect to the reference contours for the DL based contouring and two of the manual generated contour sets. For the remaining two manual generated contour sets there was a statistically significant difference (p<0.05). Statistical significance was evaluated with a Wilcoxon rank-sum test. <h3>Conclusion</h3> Deep-learning can provide reliable prostate contours from TRUS images with needles in-place. Such contours are within the variability of those provided by expert observers. The inter user variation indicates the need for more standardized contouring and the presented DL based approach could be a potential solution.