Oncology and radiation therapy

In the seven-institution evaluation, 11 physician evaluators rated 65% of contours clinically usable and 31% usable with minor edits; 4% needed major edits or redrawing. Varian Medical Systems employees were among the authors. Across 18 American Oncology Institute centers, six physicians recorded both manual contouring and editing times on 116 cases; the software saved 13.8 minutes per case on average, worth INR 37 to INR 670 net, and one author listed a Varian affiliation.

At a Delaware cancer center, in 30 patients, time savings were 67% for head and neck, 83% for pelvis and 84% for thorax, and an expert rated 50% of head and neck, 72.5% of pelvic and 82% of thoracic contours clinically usable. At a Tokyo hospital, brain, rectum and bladder contours differed significantly across software versions, and rectum accuracy worsened after an update.

Artera's predictive model drew on 5,727 men from five phase 3 trials and was validated in NRG/RTOG 9408 (1,594 men, median follow-up 14.9 years): ADT reduced distant metastasis in the 34% who were model positive but not in the 1,051 who were negative. The study was supported by National Cancer Institute grants and Artera, whose employees were authors. A long-term ADT model, validated in RTOG 9202 (1,192 men), found benefit in biomarker-positive men (subdistribution hazard ratio 0.55) but not in negative men (1.06).

In ASTuTE, 52 of 74 men (70.3%) who first chose short-term ADT dropped it after testing, while 3 of 126 (2.4%) switched to it; GenesisCare funded the single-arm trial in partnership with Artera. FDA's pivotal study of 886 patients found a 10-year metastasis risk of 28.1% in the high-risk group versus 8.1% overall. A 2025 systematic review counted 10 prostate studies of ArteraAI, none prospective.

At two BC Cancer centers, the software drafted contours for about 551 eligible central nervous system, head and neck and prostate plans from September 2019 to March 2020; in 203 surveys (32% compliance), most organ contours needed minimal edits, with mean satisfaction of 4.1 to 4.6 out of 5. Limbus employees were authors, and editing time was not measured. At Willis Knighton Cancer Center, adding the software and applicator modeling cut planning time for vaginal cuff brachytherapy by 13.7 minutes per fraction and contouring time by 6.8 minutes, though the two tools' effects could not be separated.

In an earlier test set, drafts took 0.6 minutes versus 26.6 minutes for experts on head and neck cases, with accuracy comparable to variation among experts. An independent failure analysis of 111 cases found biliary stents or drains behind four of six liver outliers and barium contrast behind 10 of 12 stomach outliers.

In the ELAISA study, run by the International Atomic Energy Agency and Aarhus University Hospital, 23 institutions were randomized to manual or AI-assisted contouring of head and neck cases. Without teaching, AI assistance raised Dice similarity with expert consensus by 0.10 for the oral cavity and 0.07 for the parotid and cut mandible contouring time by 6.60 minutes; 24% of AI contours were accepted unedited.

In a seven-institution study in Singapore and Europe, total head and neck contouring time fell from 81.9 to 44.4 minutes (46%), but lymph node levels IA, IB, III, IVA and IVB showed no significant savings, and at most clinics levels IVA and IVB took longer to edit than to draw. MVision employees were authors and handled scan randomization. In a Cypriot comparison of five systems on 80 patients, MVision's software had a median volumetric Dice of 0.88 and saved 43.6 minutes on average. Twenty French radiation oncologists found MVision's the only one of eight programs whose contours on children's CT scans met their usability threshold.