Fig. 9
From: Learning from multiple readings for axial spondyloarthritis classification of the sacroiliac joints
Example showing an SIJ as input. The immediate output \(\varvec{\hat{p}}\) is then transformed into the inter-consensus prediction via matrix multiplication with the inter-consensus annotator matrix \(A_{R1R2}\) (red). Subsequently, the prediction is then further multiplied with specific intra-reader annotator matrices \(A_{R1}\) (green) and \(A_{R2}\) (blue) resulting in two separate intra-reader specific predictions. The intra-reader predictions are then further split into session-specific readings S1, S2, and S3 for all 3 read sessions. The example shown corresponds to one prediction but in practice, we perform 14 separate predictions (3 lesion types (Oedema, Erosions, Fatty Lesions) with 4 quadrants plus 2 for Ankylosis and Sclerosis). Also, note that the input slice shown is a single 2D slice from a sequence but in actuality, we process all T1-weighted and STIR slices in combination for a single SIJ. At inference time, \(\varvec{\hat{p}_{R1R2}}\) is used as our best prediction when comparing against inter-reader labels.
