Accurate segmentation of the clinical target volume (CTV) corresponding to the prostate with or without proximal seminal vesicles is required on transrectal ultrasound (TRUS) images during prostate brachytherapy procedures. Implanted needles cause artifacts that may make this task difficult and time-consuming. Thus, previous studies have focused on the simpler problem of segmentation in the absence of needles at the cost of reduced clinical utility. To use a convolutional neural network (CNN) algorithm for segmentation of the prostatic CTV in TRUS images post-needle insertion obtained from prostate brachytherapy procedures to better meet the demands of the clinical procedure. A dataset consisting of 144 3-dimensional (3D) TRUS images with implanted metal brachytherapy needles and associated manual CTV segmentations was used for training a 2-dimensional (2D) U-Net CNN using a Dice Similarity Coefficient (DSC) loss function. These were split by patient, with 119 used for training and 25 reserved for testing. The 3D TRUS training images were resliced at radial (around the axis normal to the coronal plane) and oblique angles through the center of the 3D image, as well as axial, coronal, and sagittal planes to obtain 3689 2D TRUS images and masks for training. The network generated boundary predictions on 300 2D TRUS images obtained from reslicing each of the 25 3D TRUS images used for testing into 12 radial slices (15° apart), which were then reconstructed into 3D surfaces. Performance metrics included DSC, recall, precision, unsigned and signed volume percentage differences (VPD/sVPD), mean surface distance (MSD), and Hausdorff distance (HD). In addition, we studied whether providing algorithm-predicted boundaries to the physicians and allowing modifications increased the agreement between physicians. This was performed by providing a subset of 3D TRUS images of five patients to five physicians who segmented the CTV using clinical software and repeated this at least 1week apart. The five physicians were given the algorithm boundary predictions and allowed to modify them, and the resulting inter- and intra-physician variability was evaluated. Median DSC, recall, precision, VPD, sVPD, MSD, and HD of the 3D-reconstructed algorithm segmentations were 87.2 [84.1, 88.8]%, 89.0 [86.3, 92.4]%, 86.6 [78.5, 90.8]%, 10.3 [4.5, 18.4]%, 2.0 [-4.5, 18.4]%, 1.6 [1.2, 2.0]mm, and 6.0 [5.3, 8.0]mm, respectively. Segmentation time for a set of 12 2D radial images was 2.46 [2.44, 2.48]s. With and without U-Net starting points, the intra-physician median DSCs were 97.0 [96.3, 97.8]%, and 94.4 [92.5, 95.4]% (p<0.0001), respectively, while the inter-physician median DSCs were 94.8 [93.3, 96.8]% and 90.2 [88.7, 92.1]%, respectively (p<0.0001). The median segmentation time for physicians, with and without U-Net-generated CTV boundaries, were 257.5 [211.8, 300.0]s and 288.0 [232.0, 333.5]s, respectively (p=0.1034). Our algorithm performed at a level similar to physicians in a fraction of the time. The use of algorithm-generated boundaries as a starting point and allowing modifications reduced physician variability, although it did not significantly reduce the time compared to manual segmentations.
Read full abstract