1046 - AI Auto Contouring and Its Impact on Resident Education

A. S. Qian¹, C. Phuong¹, E. Porter¹, L. Ni¹, F. Mohebi², J. J. Chen¹, R. Sabol¹, N. V. Kotha³, J. C. Hong¹, and S. E. Braunstein¹; ¹Department of Radiation Oncology, University of California San Francisco, San Francisco, CA, ²University of California, Berkeley, Berkeley, CA, ³Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA

Purpose/Objective(s): The integration of AI auto contouring (AAC) in radiation oncology has streamlined the delineation of organs at risk (OARs). Assessing OAR contours is a vital skill in radiation oncology. This study aimed to elucidate the impact of AAC implementation on resident education, contouring skill development, and longitudinal educational programming. Materials/

Methods: An anonymized survey was administered to all residents and resident-interfacing faculty at a single tertiary academic institution that implemented AAC in the past year. Respondents indicated whether they were resident or faculty and answered questions in free text or on a Likert scale from 1-5 (strongly disagree (1) to strongly agree (5)). Free text responses were analyzed qualitatively. Statistical analyses with t-tests were performed, with p values = 0.05 considered as significant. AAC contours that residents worked on were analyzed to determine the percentage of OARs edited by residents. Contours were considered edited if an entire axial slice was added or removed, or an axial slice was edited to a Hausdorff distance of = 2mm.
Results: 13/13 residents (100% response rate) and 14/21 faculty (67% response rate) completed the survey. Residents and faculty disagreed on whether AAC improved understanding of anatomy (4.2 vs 2.1, p<0.01) respectively, or had a positive impact on general resident education (4.2 vs 2.4, p<0.01). However, both agreed that AAC allowed residents to spend less time contouring (4.6 vs 4.3, p=0.22) and improved workflow from simulation to plan approval (4.4 vs 3.9, p=0.13). Both were relatively neutral on the quality of AAC contours (3.4 vs 2.6, p=0.06). Residents and faculty also agreed that AAC improved understanding of standardized OAR nomenclature (4.1 vs 3.2, p=0.02) and positively contributed to clinic (4.8 vs 3.6, p< 0.01) and resident wellbeing (4.6 vs 3.4, p< 0.01), though at varying levels. On objective analysis of 113 imaging studies with AAC, there were 1526 edited OARs, of which 1479 (96.9%) were first edited by residents. Thematic analysis of faculty responses highlighted inaccuracies of AAC OARs, with concern that the lack of OAR contouring repetition may translate into inadequate experience with and knowledge of CT anatomy. Residents highlighted that although AACs were imperfect and require review and revision, the time saved improved quality of life and could be better channeled into other educational endeavors.
Conclusion: Residents and faculty disagree on whether AAC affects understanding of OARs, specifically highlighting the imprecision of AI auto contours. However, most agree that there has been a positive impact on clinic and overall education. Improving the integration and understanding of AAC-derived OARs during contouring will be crucial for improving resident training and ensuring high-quality care delivery in the AI era.