In-line Automated Scan Quality Assessment and Correction for Archival Histopathology Slide Scanning

Andrew P. Norgan, M.D., Ph.D.¹, Bryan J. Dangott, M.D.², Prasanth Perugupalli, M.S.³, Jason Ross³ Kurt E. Simon, M.B.A., PMP¹, Darin P. Morgan, Stephanie A. Derauf, PMP, and Thomas J. Flotte, M.D.^1
1Department of Laboratory Medicine and Pathology, Mayo Clinic, Rochester, MN; ²Department of Laboratory Medicine and Pathology, Mayo Clinic, FL; nference, Inc.,³Cambridge, MA

Methods

With a single human operator, we digitized 23,916 slides over 30 days (approximately 800 slides per day) using a research-use-only 4-head scanning system (nference, Inc). Slides were selected from the Mayo Clinic Pathology Tissue Archive and represented periods ranging from the 1950s to present. No cleaning or other preparatory steps were used to prepare slides for scanning. Using on-scanner quality models, each slide was annotated in real-time for errors detected in focus, stitching, bubbles and folds, and other detectable errors (see Figure 1). A subset of images from a quality control slide set (100 well-characterized slides) were also reviewed manually.

Detectable Features

In-line feature detection was capable of annotating and/or abstracting: 1) areas of variable tissue thickness and/or focus variation, 2) existing hand-written slide annotations, 3) tissue folds, 4) tissue fragmentation, 5) serial sections, 6) bubbles, 7)faded or unstained tissue, 8) misaligned coverslips, 9) special stains, 10) cytology preparations, 11) small or minute tissue fragments, 12) artifacts or debris, and 13) tile stitching artifacts

Feature Maps and Automated Correction

Select features, including focus and stitching, are abstracted to tile-based feature overlays. The feature overlays provide an “at a glance” snapshot of slide quality and can also be integrated into a numeric quality metric (if desired) to drive in-line or downstream quality control activities. In our testing, we used real-time on-scanner out-of-focus assessment to provide an automated trigger for autonomous rescan of the image with an increased density of focus points. While use of this quality control feature increased scanning time of out-of-focus scan slides by 1 to 2 seconds, it provided a significant improvement in 1st scan acceptability without human (manual) intervention.

High Throughput Cluster

Scanning was accomplished using a research use only cluster of 4 single slide line scanners using 40X zoom (60X optional) with motorized magnification selector at 0.26µm/pixel scan resolution. The 4 scanners are fed by an automated robotic arm from a single common slide tray. Each scanner is independent of the others and can complete scanning (or re-scanning as necessary) without disrupting the operation of its cluster partners.

Scanning Features

The scanners capture dynamic Z-stacks of each scan area and utilize Z-stacks to perform real-time focus assessment and continual refinement of the optimal focal plane. Real-time focal quality assessment allows for in-line quality control and autonomous re-scanning as needed. The scanners render final images as DICOM objects with lossy-compressed JPEG2000 pixel data.

Quality assurance example

Automated in-line error detection initiates automatic re-scan to correct most errors at a tissue level without human intervention.

Outcomes

The overall acceptability rate of scans in this pilot was 99.5%. In total, 118 slides were deemed to have unacceptable scan quality. However, this includes slides that had physical defects that would have also rendered the physical slides unacceptable for any use. This error rate is ~3-6x lower than published rescan rates using leading systems.

Focus errors were detected algorithmically in 96 (0.4%) scanned slides and post-scanning stitching errors were algorithmically detected in 34 (0.14%) of slides.

A manual review of 100 selected slides verified 100% of errors that were algorithmically detected. In addition, 2 slides without algorithmically detected errors were found to have significant focus and/or stitching errors.

Conclusions

In-line quality assessment allows for rapid recognition of quality defects in scanned archival tissue slides.

Autonomous focus assessment can ameliorate focal defects while the slide is still on the scanner, potentially saving significant resources in manual review and slide handling.

Algorithmically detected errors were verified in manual review, suggesting a high specificity in the error detection algorithms.

Rate undetected focus & stitching artifacts detected during manual review suggests potential further improvements in algorithm sensitivity are possible.

References

1. Whole slide imaging equivalency and efficiency study: experience at a large academic center. Hanna et al., 2019
2. Complete Digital Pathology for Routine Histopathology Diagnosis in a Multicenter Hospital Network. Retamero et al., 2019
3. A quantitative approach to evaluate image quality of whole slide imaging scanners. Shrestha et al., 2016