To Evaluate the Clinical Implementation Feasibility of the Siemen's Auto-Contouring Deep-Learning Solution, AI-Rad Companion Organs RT
L Maduro Bustos1,2*, L Doyle1, D Nurbagandova1, J Noonan1, A Sarkar1, K Andreou1, P Graner3, K Saddi3, F Mourtada1,2, (1) ChristianaCare, Newark, DE, (2) Thomas Jefferson University, Philadelphia, PA, (3) Siemens Healthineers, Malvern, PA.
Presentations
TU-I345-IePD-F6-6 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 6
Purpose: To evaluate the clinical implementation feasibility of the Siemens Healthineers AI-Rad Companion Organs RT (AIRCORT) contouring algorithm for prostate organs at risk.
Methods: Computed tomography (CT) datasets from 10 male pelvis cases were collected. Three human annotators generated contours using departmental standard tools, including auto-contouring features in RayStation (RS). A fourth set was obtained from AIRCORT; contouring times were recorded for efficiency comparison. A physician (expert) then assigned a score from 1 to 4 to each contour set using the following scale: 1-Must Redo, 2-Major Edits, 3-Minor Edits, 4-Clinically usable. Using the contours from the highest scoring human annotator for each case as a reference, a RS script was run to obtain a DICE Similarity Coefficient (DSC) between the reference and the set generated by AIRCORT. The time to make corrections to the AIRCORT’s sets scored between 1 and 3 was then measured and compared to the time recorded by the human annotators.
Results: The average DSC for the 10 cases are: bladder and left-femoral-head= 0.93, right-femoral-head= 0.91, rectum= 0.8. The expert deemed three of the auto-contoured cases as “1-Must Redo” as at least 2 structures required to be entirely recontoured. The other seven cases received a score of 2 or 3, requiring major/minor edits. The average rectification time for all was (27.9±5.7)min.
Conclusion: AIRCORT improves the overall contouring efficiency by 35.4%. The mean DSC for all structures proves AIRCORT to be useful as a starting point for the human annotators. The rectum average DSC was low (0.8) due to the selection of the sigmoid/rectum interface determined by AIRCORT; the software includes about 2-3cm of what is considered the sigmoid in our clinical practice. Future evaluation will focus on extending this work to additional disease sites and identifying systematic improvements for low-performing OARs.
Funding Support, Disclosures, and Conflict of Interest: Evaluation study of AI-rad Companion Organs RT is supported in part by Siemens Healthenieers Inc.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email