Session Item

It is well established that patients with cancers of the head and neck (H&N) are adversely affected by treatment delays. Availability of qualified clinicians is often the rate-limiting step in the treatment planning pathway, and manual contouring of targets and organs-at-risk (OARs) is the most time-intensive stage for clinicians. Novel AI-based auto-contouring algorithms are becoming commercially available. We aimed to test whether using Mirada DLCExpert (Mirada-Medical, Oxford, UK) with a H&N model trained at another institution could improve clinical contouring efficiency for OARs.

DLCExpert was used to generate spinal cord, brainstem and parotid volumes for 10 H&N patients previously contoured by experienced clinical oncologists at our institution. The quality of the automatically-generated contours was assessed by: quantitative measurement of congruence with existing clinical contours (centre-of-mass shift; dice similarity co-efficient (DSC)); blinded review of both existing clinical contours and DLCExpert contours, scored 1-5 based on level of modification required (1=complete; 2=major; 3=minor; 4=insignificant; 5=none). Time taken to generate clinically acceptable contours for all OARs routinely outlined in our institution (those listed above, plus orbits, lenses, optic nerves and chiasm) was prospectively recorded for 9 H&N patients without use of Mirada, and for 7 patients for whom auto-contouring was used (DLCExpert for OARs listed above, and Mirada Embrace atlas-based software for the optical structures, for which the DLCExpert model had not been trained). All contouring, reviewing, editing and scoring of contours was carried out by a consultant clinical oncologist with extensive H&N experience.

Mean centre-of-mass shifts were close to zero, indicating no systematic spatial shift of DLCExpert contours with respect to clinician contours. DSC averaged 0.80 across all structures (see table 1).

Table 1: Mean centre of mass shift and DSC for the 10 test patients.

In the blind evaluation, the scores for clinical versus DLCExpert contours were: 3.9 vs 2.6 for Brain stem; 4.7 vs 3.6 for Spinal cord; 4.4 vs 3.6 for left parotid; 4.8 vs 3.5 for right parotid. Overall, 33% of structures were clinically acceptable with no amendments, with 50% requiring minor and 18% major adjustments.
Figure 1 shows the proportion of contours requiring major amendment for each OAR, versus those requiring minor or no adjustment.

Without the use of Mirada software, OAR contouring took an average of 10.4 minutes (range 8-18 minutes); with auto-contouring, review and amendment of OAR contours took 8.4 minutes (range 8-9 minutes).

Whilst the automatically generated contours were scored inferior to the clinicians own contours, over 80% of contours required no more than minor amendments, with 33% requiring no adjustment. This led to a reduction of approximately 20% in the time taken for a clinician to contour OARs for a H&N patient.