Session Item

Online adaptive radiotherapy (O-ART) is being introduced into clinical practice. For O-ART, treatment planning should be replanned using daily verification image because of the anatomical information change. Therefore, routine usage is limited due to the time-consuming task of manual segmentation and its accuracy variation among observers. This study proposes patch-wised (PW)-U-net-based automatic pelvic CT segmentation models for prostate radiotherapy planning and validate its performance compared with the conventional non-PW U-net model (C-U-net).

One-hundred patients who underwent radiotherapy for locally advanced prostate cancer were enrolled in this study. For all patients, CT scan was performed whole pelvis. The contours of organs including prostate, bladder and rectum delineated by the radiation oncologist for radiotherapy planning were registered as the ground truth segmentations. The dataset was split into training (n = 70) and test (n = 30) subsets. Figure 1 shows the workflow of this study. For the PW-U-net, the input CT and the corresponding bladder segmentation maps with 256 x 256 pixels were divided into 4 and 16 patches with 128 x 128 and 64 x 64 pixels, respectively. The prostate and rectum segmentation maps with 128 x 128 pixels were divided into 4 and 16 patches with 64 x 64 and 32 x 32 pixels, respectively. For each data, 2D-U-net model was constructed using the training data and, then verified with the test data. Finally, the divided images of predicted segmentation were concatenated to original image . To compare the segmentation performance between the 4-PW-U-net, 16-PW-U-net and C-U-net, dice similarity coefficient (DSC) and Hausdorff distance (HD) between predicted and ground truth segmentation for prostate, bladder and rectum were evaluated.

Table1 shows the results of segmentation performance comparisons between proposed PW-U-net and the C-U-net for each organ of all test data. The DSC for the 4-PW-U-net, 16-PW-U-net and C-U-net between predicted and ground truth segmentation for prostate were 0.84 ± 0.04, 0.66 ± 0.07 and 0.83 ± 0.04, those for bladder were 0.89 ± 0.03, 0.90 ± 0.02 and 0.86 ± 0.03, those for rectum were 0.79 ± 0.03, 0.65 ± 0.02 and 0.75 ± 0.03, respectively. The HD for prostate were 2.57 ± 0.80, 2.89 ± 0.74 and 2.57 ± 0.81 mm, those for bladder were 2.89 ± 0.91, 2.88 ± 0.55 and 2.95 ± 0.90 mm, those for rectum were 2.16 ± 0.61, 2.46 ± 0.75, 2.26 ± 0.55 mm, respectively. Although the 4-PW-U-net improved the segmentation performance of all organs over the C-U-net, 16-PW-U-net for prostate and rectum did not.

In this study, our proposed 4-PW-U-net was demonstrated to improve the performance of the pelvic CT segmentation over the C-U-net. However, our proposed method requires further studies to optimize the patch segmentation and training method for each organ and evaluate the impact of segmentation accuracy on dose determination for both the organ at risk and the target volume.