manual correction mode

Not sure this is doable or not. What I would like to suggest is to have a 'manual mode segmentation' that can overrule 'predicted segmentation'. Currently, predicted segmentation is based on the pattern that we annotate. For example, if I annotate the black region as bacteria and gray region as background, the predicted segmentation would follow that pattern. However, there are quite a number of bacteria (with non ideal black color, therefore more similar to background), and they are predicted as background (i.e. undersegmentation of bacteria). When I correct these wrongly predicted bacteria (as background) by re-annotation, many background regions (real background) are then predicted as bacteria (i.e. oversegmentation of bacteria). This is becasue I re-scribble not very black regions as bacteria, so prediction follow that pattern. And then if I try to correct this again, it goes back to the first scenario of undersegmentation. So, i end up with either over/under segmentation. This issue could happen in most dataset as the intensity/pattern of individual objects vary much even they are same class. So, if there is a manual segmentation mode that can overrule trained-prediction, I thought I can get a more accurate segmentation. For example, I can start with under/over segmentation, then I use the manual mode for correction, which is, if i select a region as bacteria in manual mode, it will segment as bacteria (even though the pattern looks more like background). Whatever I segment in manual mode should overrule the trained-prediction. So, it is kind of like auto-segmentation that allows manual correction.