Interobserver agreement of radiologists assessing the response of rectal cancers to preoperative chemoradiation using the MRI tumour regression grading (mrTRG)

•Inter-observer agreement of radiologists was assessed using MRI rectal tumour regression scale.•Kappa agreement had a median value of 0.57 (95% CI: 0.37–0.77) indicating an overall good agreement.•In 65.9% and 90% of cases the radiologists were able to correctly highlight good and poor responders respectively.•With minimal training good agreement and differentiation between good and intermediate/poor responders can be achieved.•Focus should be on facilitating the identification of good responders.

AimTo investigate whether the magnetic resonance imaging (MRI) tumour regression grading (mrTRG) scale can be taught effectively resulting in a clinically reasonable interobserver agreement (>0.4; moderate to near perfect agreement).Materials and methodsThis study examines the interobserver agreement of mrTRG, between 35 radiologists and a central reviewer. Two workshops were organised for radiologists to assess regression of rectal cancers on MRI staging scans. A range of mrTRGs on 12 patient scans were used for assessment.ResultsKappa agreement ranged from 0.14–0.82 with a median value of 0.57 (95% CI: 0.37–0.77) indicating good overall agreement. Eight (26%) radiologists had very good/near perfect agreement (κ>0.8). Six (19%) radiologists had good agreement (0.8≥κ>0.6) and a further 12 (39%) had moderate agreement (0.6≥κ>0.4). Five (16%) radiologists had a fair agreement (0.4≥κ>0.2) and two had poor agreement (0.2>κ). There was a tendency towards good agreement (skewness: 0.92). In 65.9% and 90% of cases the radiologists were able to correctly highlight good and poor responders, respectively.ConclusionsThe assessment of the response of rectal cancers to chemoradiation therapy may be performed effectively using mrTRG. Radiologists can be taught the mrTRG scale. Even with minimal training, good agreement with the central reviewer along with effective differentiation between good and intermediate/poor responders can be achieved. Focus should be on facilitating the identification of good responders. It is predicted that with more intensive interactive case-based learning a κ>0.8 is likely to be achieved. Testing and retesting is recommended.