Fig. B.2.

Extended example of Fig. 2M-P showing poor agreement in the candidate segmentation failures of a T1-map affected by an extracardiac structure (breast implant). The top left image shows the manual segmentation and the top right image shows the visualization of the candidate segmentations. The rest shows the candidate segmentations from U-net 7 to U-net 27, and the combined segmentations (Model 1 - Model 6). When the candidate U-nets failed, they appeared to fail differently, as demonstrated by U-net 7, 11, 23, and 27, obtaining a DSC of 0.59, 0.68, 0.54, and 0.49, respectively. U-nets 7 and 11 failed to form an annulus-like myocardial mask, whereas U-nets 23 and 27 falsely identified the breast implant as part of the myocardium. Despite the difficulty, U-nets 15 and 19, and Combined Models 3 and 4 successfully segmented the myocardium, with DSCs $\ge 0.77$. This illustrates the importance of including a diverse set of candidate models. Combined Model 3 (in the red box) achieved the highest predicted DSC (0.80), thus its segmentation was selected as the final output by the QCD framework. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)