ESTRO 2024 - Abstract Book

S3127

Physics - Autosegmentation

ESTRO 2024

significant differences emphasising the need for review of every patient case. Table 1 shows the surface DICE scores from the evaluation.

From the clinical audit, Table 1 shows the surface DICE and the percentage of OARs manual adjusted from 0-3 and 3 6 months.

Some OARs showed differences in surface DICE values from evaluation to clinical implementation due to the fact that it is not always necessary to manually contour the full superior-inferior extent of an OAR that only has maximum dose constraints applied.

Some changes in user behaviour are apparent by comparing the results at each time point.

• 3 months after clinical implementation some changes in practice were made to contouring for femoral heads and spinal cord which are reflected in the results for these structures. • A reduction in editing of optical structures may be due to our previous local practice being to contour a more conservative structure than the guidelines, and Drs gradually becoming more comfortable with using the guideline based structure. • Other OARs are highlighted for further investigation – where there has been a reduction in contour editing, this could be due to automation bias or due to users becoming more comfortable with not making adjustments which will have negligible clinical impact. Table 1 Comparison of DICE scores from the evaluation (manual contour compared to deep learning) to clinical implementation (deep learning contour compared to manually adjusted deep learning contour), and the % number of manual edits for each OAR. The surface DICE value is the average from the manually edited contours (excludes unedited contours).