Head-To-Head Performance Comparison of Two Deep Learning Segmentation Algorithms for Radiotherapy Planning: A Study in Prostate
H Martinez1*, B Rich2, L Young3, F Yang2, (1) Department Of Biomedical Engineering, University Of Miami, Coral Gables, FL (2) Department Of Radiation Oncology, University Of Miami, Miami, FL (3) Department Of Radiation Oncology, University of Washington, Seattle, WA
Presentations
PO-GePV-M-116 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Deep learning (DL) based automated segmentation methods have demonstrated the potential to improve the radiotherapy (RT) treatment planning process by offering readily available, while consistent, anatomical contours. There is, however, a lack of knowledge of how reliable and valid these algorithms are for clinical use. This study aimed to evaluate performances of two marketed DL segmentation algorithms in a head-to-head study design.
Methods: The algorithms being assessed included ProtégéAI (MIM Software Inc., Cleveland, OH) and AccuContour (Manteia Tech, Xiamen, China). Imaging data used consisted of CT simulation scans of 100 prostate cancer patients. The segmentation accuracy of each algorithm for prostate, bladder, femoral heads, rectum, and seminal vesicles was compared to the respective physician contour using metrics of Dice coefficient (DICE) and absolute volumetric difference (AVD) with differences being identified.
Results: AccuContour significantly outperformed ProtégéAI in delineating bladder and right femoral head with mean DICE of 0.88 Vs. 0.74 (two-sided paired t-test; p<0.001) and 0.87 Vs. 0.82 (p<0.001) and mean AVD of 0.14 Vs. 0.25 (p=0.007) and 0.18 Vs. 0.36 (p<0.001); whereas the later excelled in delineating prostate and rectum with mean DICE of 0.83 Vs. 0.67 (p<0.001) and 0.83 Vs. 0.78 (p<0.001) and mean AVD of 0.17 Vs. 0.36 (p<0.001) and 0.13 Vs. 0.18 (p=0.0149). For seminal vesicle, ProtégéAI showed superior accuracy in terms of DICE (mean: 0.65 vs. 0.49; p<0.001) while no significant difference according to AVD (mean: 0.44 vs 0.54; p=0.348) in comparison to AccuContour. Regarding left femoral head, no significant difference was seen between the two algorithms regarding DICE (mean: 0.91 vs 0.90; p=0.598) or AVD (mean: 0.13 vs 0.10; p=0.218).
Conclusion: Accuracy of DL-based segmentation varied based on algorithm and with anatomical location under evaluation. The findings from the current study indicate that caution should be exerted when implementing these algorithms into clinical workflow.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email