The IHC project I worked on at A*STAR involved finding tissue in pathology slides, separating stains, segmenting nuclei, extracting features and classifying them.
Nuclear segmentation needs close inspection. Touching nuclei can be merged, and boundaries can be misplaced. The segmented regions are then used to extract features, so those errors can affect classification as well.
Comparing the original and segmented images helps reveal problems that the final accuracy score alone won’t explain. The classifier receives extracted features; it has no way to check whether the original nucleus was segmented correctly.
I’ve put the methods, images and evaluation results from my report on the IHC project page.