Purpose: Applications of automation are increasingly expanding into different aspects of cancer management in radiation therapy. At the forefront is the contouring of organs at risk utilizing artificial intelligence. Few studies, however, have utilized artificial intelligence for target segmentation in the pelvic region. We implement a first of its kind deep learning contouring (DLC) for the prostate and nodes clinical target volume (CTV). This study investigates the similarity of manual delineated CTV and planning target volumes (PTV), done by radiation oncologists, with ones completed using autosegmentation with a DLC model.
Methods: DLC model training used one hundred patient’s contours of the prostate and nodes. The model was validated by generated CTV contours on 30 patients and expanding by 0.5cm to get PTV. Fifteen of those patients had physician CTV contours and subsequent expansion for PTVEX, and fifteen had direct PTV (PTVDIR) volumes by the physician. Evaluation of the geometric similarity using the Dice Similarity Coefficient (DSC) and average distance (AD), a modification of the Hausdorff Distance, allows for contour comparison.
Results: Average values and standard deviation for the DSC generated on the CTV, PTVEX, and PTVDir are 0.68±0.08, 0.77±0.05, and 0.75±0.05. AD yielded 5.5±1.1m, 6.0±1.2mm, and 6.5±1.2mm. Typically, deviations occurred at the superior aspect of the nodal volume and on the prostate’s inferior and posterior, attributed to known inconsistencies in the training volume. A subjective observation by physicians yields that, in general, the contours are beneficial and should yield time savings.
Conclusion: We demonstrate the automating DLC for the prostate and nodes is possible. Although not statistically significant, we find an increased correlation between PTV volumes when physician contours the CTV over PTV. Further curating of training data should yield increased effectiveness in CTV contouring.
Funding Support, Disclosures, and Conflict of Interest: Research performed in collaboration with Mirada Medical
Not Applicable / None Entered.