Purpose: To develop and clinically review a deep learning segmentation algorithm that segments intact and post-operative targets (prostate, prostate bed, seminal vesicles (SV), and SV fossa) and five normal structures for prostate cancer treatment planning.
Methods: A modified two-stage 3D UNet model was developed using retrospective clinical CT data from 670 patients (8:1:1 train, validation, test split). Stage one inputs a downsampled CT volume and outputs a localized coarse region of interest (ROI) containing multiclass segmentations. The ROI segmentations from stage one are then processed using k-means clustering to identify centroids for 3D patches that are cropped from the full-resolution CT image to generate final segmentation predictions. The 3D patches are then combined by averaging pixel-wise prediction probabilities when there is overlap between 3D patches to generate individual ROI segmentations. The resulting tissue contours were then evaluated quantitatively (Dice, mean surface distance) and qualitatively (radiation oncologist review).
Results: Segmentations of 9 planning structures were generated with the following average Dice/mean surface distances: prostate 0.79±0.11/2.9±2.0mm, prostate bed 0.75±0.05/3.1±0.6mm, SV 0.6±0.17/3.7±4.3mm, SV-Fossa 0.49±0.18/4.2±3.1mm, rectum 0.87±0.05/2.1±1.0mm, bladder 0.96±0.01/0.8±0.1mm, penile bulb 0.74±0.11/1.6±0.8mm, femoral heads 0.92±0.03/1.6±1.0mm, sigmoid 0.40±0.25/12.6±13.5mm. Two radiation oncologists each reviewed at least two intact and two post-operative patient datasets. The preliminary qualitative review of contours showed the majority of bladder, penile bulb, and femoral head contours were acceptable without edits; the rectum, prostate bed, and SV contours required minor edits; and the sigmoid, SV fossa, and prostate contours required either minor or major edits.
Conclusion: We developed and evaluated an auto-segmentation algorithm for intact and post-operative prostate radiotherapy. Additional qualitative review is necessary to determine clinical acceptability of the automated contours.
Funding Support, Disclosures, and Conflict of Interest: This material is based upon work supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. 2043424. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.