ESTRO meets Asia 2024 - Abstract Book

S396

RTT – Treatment planning, OAR and target definitions

ESTRO meets Asia 2024

Artificial Intelligence (AI) presents numerous opportunities for improving efficiency and consistency of practice within radiation oncology. For these benefits to be leveraged clinicians must have confidence in the output of AI systems. The integration of such systems into pre-existing clinical workflows and staffing models must also be actively facilitated. This mixed-methods retrospective cohort study evaluated initial clinician experience of an AI enhanced workflow for breast target volume delineation.

Material/Methods:

Clinicians with breast cancer expertise from a single centre were invited to participate in the study. 15 previously treated, anonymised, breast cancer patient datasets were assigned to each participant. Structures from each dataset were removed and a commercially available, deep learning AI application was used to generate a clinical target volume (CTV_AI) for 10 whole-breast (WB) and 5 axillary nodal levels I-IV (SCAX) cases per participant. Participants scored compliance of each CTV_AI with contouring guidelines used to train the AI models, clinical acceptability of the unmodified CTV_AI, and perceived time-saving using the CTV_AI to generate a clinically acceptable CTV (CTV_AI-mod). Strongly agree, agree or neutral responses were grouped for reporting. Absolute volume, sensitivity, precision, Dice similarity coefficient (DSC) and mean Hausdorff Distance (mHD) were calculated for each case comparing CTV_AI-mod, CTV_AI, and the originally treated CTV (CTV_clinical). Each participant took part in a semi-structured interview to explore their perceptions utilising the AI-enhanced workflow. Qualitative analysis was performed on interview transcripts using directed content analysis to establish themes consistently expressed. This study was approved by the University of Otago Human Research Ethics Committee (ref HD23/069). Three of four clinicians elected to participate in the study. 37 of 45 assigned datasets were completed (25 WB, 12 SCAX). Consistency of CTV_AI structures with contouring guidelines was reported for 84% and 83% of WB and SCAX cases, respectively. Clinical acceptability of CTV_AIs was most commonly rated as requiring no edits or minor edits only for both WB and SCAX cases (88% and 75%, respectively). Based on early feedback, a protocol amendment was introduced to provide additional de-interpolated versions of each CTV_AI structure, and one participant voluntarily repeated 5 cases completed prior to the amendment. Of cases completed following protocol amendment, time-saving due to availability of a CTV_AI was reported in 86% of cases (92% and 78% of WB and SCAX cases, respectively). Contour similarity metrics are reported in Table 1. High similarity between CTV_AI-mod and CTV_AI structures suggested a limited extent of required clinician editing. CTV_AI-mod structures were generally observed to be smaller compared to CTV_AI structures. Higher precision relative to sensitivity for CTV_AI-mod/CTV_AI comparisons indicated that the AI platform more commonly over-contoured than under contoured relative to clinician modifications. Four themes were established from analysis of clinician semi-structured interviews (Table 2). Overarchingly, participants held a positive view of AI integration within the clinical setting. They valued the opportunity to engage critically in the process of workflow development that this study afforded them, even where challenges or barriers were identified. Motivation to implement the refined workflow and explore its benefits across wider treatment indications was consistently expressed. Results: