Purpose: To quantitatively compare structures contoured using a deep learning model to those that were manually contoured on prostate radiotherapy planning CT images. Contours of the prostate, bladder, and rectum were investigated.
Methods: We randomly selected 100 CT image sets of prostate radiotherapy patients containing expertly contoured and peer-reviewed structures representing the prostate, bladder, and rectum. For each image set, we also generated contours using the Deep Learning Male Pelvis model in Elektaâ€™s Advanced Medical Image Registration (ADMIRE) software (v3.0). This software is currently available only for research purposes, and remains under active development. The deep learning model in this software was created using training data from other institutions that were entirely independent from the testing data used in this study. Agreement between contours was evaluated using the dice similarity coefficient and average Hausdorff distance.
Results: Dice similarity coefficients (mean Â± standard deviation) of 0.86 Â± 0.04, 0.96 Â± 0.02, and 0.84 Â± 0.06 were achieved for the prostate, bladder, and rectum, respectively. Average Hausdorff distances for the same structures were 2.0 Â± 1.9 mm, 0.9 Â± 0.5 mm, and 2.4 Â± 1.9 mm. The rectum results in particular were skewed by outliers, which resulted from inconsistent local practice in manual delineation of the superior edge of the rectal volume. The median dice score for the rectum was 0.86, and the median average Hausdorff distance was 1.7 mm.
Conclusion: ADMIRESâ€™s Deep Learning Male Pelvis model provided contours of the prostate, bladder, and rectum that agreed well with an independent set of expertly delineated contours. Future work will investigate and quantify the improvements gained by using data from our institution in the model training, in an effort to reduce the effects of inter-institutional variations in contouring practices.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by Elekta.