Figure 4
From: Automated quality assessment of large digitised histology cohorts by artificial intelligence
(A) QA of our annotated patch dataset (from ProMPT and contemporary cohorts) using other quality tools. (A) Comparison of performances of HistoQC, our proposed set of hand-crafted features, and PathProfiler for usability and artefact classification. (B) Examples of sample artefacts that are challenging to identify using hand-crafted features; (a, b) staining issues cannot be generally related to the brightness of images, (c, d) ‘other’ artefacts such as calcification (c) and dirt (d) that cause regions of low intensity variation are confused with out-of-focus regions. This may falsely reinforce recommendation by the algorithm for a slide re-scan, although this will not resolve those artefacts, (e, f) simple hand-crafted features such as Laplacian filtering misclassify unusable regions that contain rapid intensity changes as usable, (g, h) hand crafted features mostly associate folded tissue with darker colours in the data and therefore cannot detect folded areas within the range of average data colours, and (i, j) thicker tissue in a section may be misclassified as folded tissue.
