ROECSG report: Real-world implementation patterns and perceptions of artificial intelligence in radiation oncology

While international surveys have examined radiation oncology professionals' attitudes toward artificial intelligence (AI) and documented generally low early utilization, actual implementation patterns and the depth of integration among engaged practitioners remain unclear. This multi-institutional survey aims to characterize current AI implementation, identify critical barriers to optimization, and evaluate the psychological factors driving clinical trust.

A de-identified survey covering six domains was developed by a multi-institutional working group and distributed via a widely subscribed medical physics listserv and at the 2025 American Association of Physicists in Medicine (AAPM) Annual Meeting. The instrument assessed adoption rates, specific workflow applications, and user perceptions (trust and understanding) using 6-point Likert scales to reduce central tendency bias. Quantitative data were analyzed using Pearson correlation coefficients and Welch's two-sample t-tests, while qualitative responses underwent inductive thematic analysis (κ = 0.78).

In this convenience sample of 80 respondents (46.3% radiation oncologists, 37.5% medical physicists), 85.0% (68/80) reported current AI use. Adoption is heavily concentrated in auto-contouring (88.2% of AI users), whereas diagnosis/prognostication (14.7%) and treatment planning (23.5%) remain limited. Most users (60.3%) employ AI for task automation without influencing clinical judgment. Trust in AI (3.55 ± 1.02) correlates positively with perceived time savings (r = 0.57), clarity of use (r = 0.54), and technical understanding (r = 0.51, all p < 0.001), with no significant difference among these correlations (Williams test, both p > 0.59). Notably, self-rated understanding averaged near the scale midpoint (3.44 ± 1.22), yet 88.8% welcomed greater AI integration. Qualitative analysis identified transparency (38%) and validation protocols (31%) as the primary barriers to deeper implementation.

AI use was common within this sample (85.0%), although self-selection likely inflates this figure relative to the broader field. However, implementation remains "broad but shallow," focused on workflow automation rather than clinical intelligence. The gap between enthusiasm and understanding necessitates a shift toward active institutional governance and standardized AI-literacy education.