Multi-Year Clinical Experience with In-House Developed AI Auto-Segmentation for Radiotherapy Planning
S Elguindi1*, J Jiang1, A Apte1, A Iyer1, E LoCastro1, Y Hu1, E Cha2, E Gillespie2, I Onochie2, D Gorovets2, M Zelefsky2, S Berry1, M Thor1, J Deasy1, L Cervino1, H Veeraraghavan1, (1) Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States (2) Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, United States
Presentations
TH-E-TRACK 4-2 (Thursday, 7/29/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]
Purpose: To report on the clinical utility of in-house developed deep learning auto-segmentation (DLA) in radiotherapy at our institution.
Methods: Two DLAs were developed to segment OAR and target volumes in routine prostate and head and neck (HN) radiation therapy. Output contours included bladder, rectum, prostate and seminal vesicles, penile bulb, left/right parotid (LP, RP), left/right submandibular glands (LS, RS), brainstem, and mandible. Both use different network architectures, a DeepLabV3+ pre-trained on ImageNet for prostate, and a Nested-Block Self-Attention method for HN. Training data consisted of 50 and 48 internally curated datasets for prostate and HN respectively. Each DLA is auto-triggered through our software pipeline, Ensemble of Voxel-wise Attributers (EVA), which manages DICOM transfers to our clinically approved high-performance computing cluster, input pre-processing, and RTSTRUCT generation. Clinical utility was gauged retrospectively through periodic random audits of comparison between the DLA output and final contours using published metrics Added Path Length (APL) and Surface DSC (SDSC).
Results: To date our DLA have segmented 1277 prostate patients (314/25% audited) since June 2019 and 337 HN patients (79/23% audited) since May 2020. Mean APL (cm) and SDSC for each structure were bladder [107.9cm, 0.96], rectum [51.2cm, 0.92], prostate and seminal vesicles [129.1cm, 0.82], penile bulb [7.1cm, 0.94], LP [39.4cm, 0.86], RP [42.9cm, 0.85], LS [22.1cm, 0.82], RS [20.2cm, 0.83], brainstem [28.1cm, 0.88], and mandible [25.1cm, 0.94]. Average total manual contouring for all structures was reduced from 2586 cm to 473 cm, representing an average reduction of 82% across both DLAs.
Conclusion: We successfully integrated two unique DLAs into routine clinical practice. This resulted in >80% reduction in APL, which has been shown to correlate with time savings. High contour surface similarity (>0.80) was achieved in all structures. This supports potential for artificial intelligence to improve the clinical workflow for segmentation in radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: Erin Gillespie is a co-founder of the website eContour.org. Joseph Deasy is a co-founder/shareholder of PAIGE.AI. Michael Zelefsky serves as a consultant for Boston Scientific. Sean Berry/Harini Veeraraghavan held research grants with Varian Medical Systems. This work was partially supported by MSK support grant/core grant P30 CA008748/NCI 5 R01 CA198121-04
Handouts
Keywords
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email