ESTRO 2024 - Abstract Book

S3057

Physics - Autosegmentation

ESTRO 2024

Purpose/Objective:

Manual contouring of OARs and CTVs is a time consuming process and often a bottleneck in the radiotherapy planning pathway. NICE have recommended AI-autocontouring technologies be implemented into clinical use as an aid to contouring for radiotherapy treatment planning, and that further evidence is collected to continue to validate these emerging technologies (1). Expert clinical review and modification of contours remains an essential step.

We aimed to assess whether Limbus AI software can be safely implemented into the radical lung radiotherapy pathway for the contouring of relevant OARs.

Material/Methods:

Prior to clinical implementation, LIMBUS AI contours (Limbus v1.7) were compared with manual contours outlined by a Consultant Clinical Oncologist for OARs relevant to radical lung radiotherapy: lungs, spinal cord, oesophagus, heart +/- brachial plexus. The brachial plexus was only contoured if it was close to the PTV. This was done for 10 previously treated patients via visual assessment and using the Dice similarity coefficient (DSC). DSC represents the overlap of volumes with a value of 0 indicating no overlap and 1 a perfect overlap. Values >0.7 suggest excellent agreement. Following the clinical implementation of Limbus an audit was performed on the first 20 lung patients. This studied the geometric similarity, using DSC, between the final clinical contour (Limbus + modification) and the unedited Limbus contour. The degree of editing required by the clinician for each Limbus contour was also assessed using a scoring system (table 1).

To further assess risks, the clinical radiotherapy plan was copied onto unedited Limbus contours and OAR doses were compared at the level of departmental tolerance dose limits.

Table 1. Scoring of Limbus contours

Results:

Limbus AI compared well with manual contours (median DSC > 0.7) for all OARs apart from the brachial plexus. Visual assessment showed a consistent difference in the superior extent of the heart contoured, but otherwise, generally contours agreed well. Limbus AI performed less well for patients with atypical anatomy: excessive stomach gas was outlined as lung in one patient and contours were all poor for a patient with a right pneumonectomy.