用于数字化和标准化存档泌尿生殖系统癌症切片的高通量切片扫描流程

Objective: To build a reproducible digital slide archival pipeline and tissue image repository for genitourinary cancers using archival slides from a large cancer research organization for digital pathology applications and biomarker discovery. Background: Digital pathology applies computer vision to digitized H&E to quantify tumor microenvironment and architectural features. Archival tissue with extensive clinical follow-up is essential to this aim, but converting glass slides to high-quality digital images is a significant challenge. We describe a high-throughput, whole slide imaging (WSI) pipeline developed at HistoWiz and its application to over 20,000 legacy cancer slides from the Gelb Center for Translational Research at the Dana-Farber Cancer Institute (DFCI). Methods: HistoWiz designed a workflow for (i) standardized intake of archival clinical slides with heterogeneous stain quality, labeling formats, and coverslipping artifacts, (ii) proprietary bulk logistics solutions enabling transport of up to 9,600 slides per shipment with <0.001% damage rate, (iii) high-fidelity scanning using hardware capable of penetrating dirt/film layers commonly found on aging slides, (iv) a real-time, scalable review process via the HistoWiz PathologyMap platform and (v) optical character recognition (OCR) for handwritten slide labels. Results: Archived slides from over 20 years (2001 to 2025) of slide collection were scanned at 900 slides per day, using two high-throughput scanning clusters with a >90% initial QC pass rate. A representative subset of images was independently evaluated by a pathologist for issues including tissue folds, cracking, air bubbles, blebs, ink annotations, and out-of-focus regions. We found that most QC failures were caused by four recurring issues: tissue folding from microtomy (30%), cover-slip issues (10%), mounting media residue (2%), and handwritten annotations or processing artifacts (1.5-12.5%). Corrective measures, including rescanning, xylene- or alcohol-based coverslip cleaning, and recoverslipping, resolved >95% of cases with cover-slip or mounting media issues, with limited tissue loss. Final slides were stored as pyramidized OME-TIFF files at 0.248 µm/pixel, averaging 2 GB per slide. In an offline Python pipeline, a Google Vision-derived OCR model yielded superior barcode read rate and accuracy on histology slides, outperforming Microsoft's OCR, Tesseract (v4+), and a Keras-based CRNN baseline. Conclusion: Our large-volume and high-throughput WSI archival workflow delivers a scalable imaging pipeline that digitized a 20-year academic tissue biobank and provides on-demand access to AI-ready, high-quality quality full-resolution slide images. This digital image repository will support large-scale digital pathology research to discover biomarkers in genitourinary cancer.