Figure 5
From: Automated quality assessment of large digitised histology cohorts by artificial intelligence
(A) H&E images demonstrating commonly encountered artefacts affecting digital WSI quality. Image quality may be affected by multiple artefacts, or just one, and focus quality can be specifically related to scanning/focus issues or it can be a result of other artefacts. (a, b) focus issue, (c) fading/loss of contrast of H&E stain, (d, e) both staining and focus issues (f, g) a combination of focus and staining issues with tissue folding and dirt, (h) bubble under the coverslip and slight focus issues in visible tissue area, (i) edge of the coverslip (affecting focus), (j) coverslip glue. (B) PathProfiler QA pipeline. After tissue segmentation, patches of \(256 \times 256\) are extracted at \(5\times\) magnification and resized to \(224 \times 224\) to accommodate for the ResNet18 CNN model. For each patch, the trained model predicts the presence of an artefact, and for focus and H&E staining artefacts it also predicts a quality score for each patch, 0 = no quality issue, 0.5 = slight quality issue, 1.0 = severe quality issue. A quality overlay is generated for each output category. In the next step, we map the predicted quality overlays to the slide-level standardised scoring system. For this, statistical parameters of quality overlays are used to predict slide-level quality scores; overall usability of the WSI (binary 0 or 1), and a score 0–10 for quality of focus and H&E staining from the lowest quality to highest quality, where the cut-off score for acceptable quality for diagnostic purposes is 4. (C) The composition of the pathologist-annotated image patches extracted from selected WSIs of ProMPT and contemporary cohorts (combined) (D) The distribution of pathologist-annotated (reference) quality scores of WSIs selected from ProMPT and contemporary cohorts (combined) for WSI usability (binary 0 or 1), focus and H&E staining quality (0–10, as above).
