使用深度学习对肺癌全切片图像进行临床级质量控制和染色协调增强

Background: Deep learning models trained on hematoxylin-eosin (H&E) whole-slide images (WSIs) increasingly support prognostic and therapeutic-response biomarkers in lung cancer. However, their reliability is often compromised by image blur, scanner and pen artifacts, and inter-slide stain variability that obscure tissue morphology and introduce spurious predictive signals. To mitigate these issues, we developed a comprehensive, quantitative WSI quality-control (QC) and stain-harmonization pipeline tailored to single-pattern lung adenocarcinoma and assessed its downstream impact on clinical outcome prediction. Methods: We analyzed 143 H&E WSIs (20×, 0.5 μm/pixel) from a Dartmouth lung adenocarcinoma cohort. QC incorporated multiple slide-level metrics, including tissue masks excluding artifacts, quantification of local blur using variance of the Laplacian across ≤600 tissue-centered crops, thumbnail-derived brightness statistics, and hematoxylin optical-density medians. Slides were retained only if they satisfied strict thresholds: tissue coverage ≥40%, artifact fraction ≤1%, blurry-tissue fraction ≤5%, brightness 140-210, and hematoxylin median 0.10-0.35. Among QC-cleared slides, an automated procedure selected a cohort-representative reference slide based on proximity to median brightness and hematoxylin metrics; its Macenko stain vectors and 99th-percentile concentration parameters were used for harmonization. All accepted WSIs were normalized to this reference. From normalized slides, 256×256 tiles (stride 256) with tissue fraction ≥70% and artifact fraction ≤2% were extracted to train a slide-level binary outcome model. Results: Of 143 WSIs, 140 (97.9%) passed QC; three were excluded. Retained slides had a median resolution of 33,792 × 46,080 pixels. From 128 QC-passing cases, we obtained 215,220 analysis-ready tiles (median 1,590 per slide). By comparison, a simpler Otsu-based pipeline produced 249,073 tiles; thus, QC-aware masking removed 13.6% of candidates-primarily low-tissue or artifact-laden regions-without diminishing tumor or stromal coverage. Stain-normalized previews demonstrated highly consistent H&E appearance. Classifier performance improved from 90.63% ± 9.36 without QC/harmonization to 95.32% ± 4.05 following implementation. Conclusions: This standardized QC and stain-harmonization framework effectively excludes low-quality slides and artifact-heavy regions while substantially improving the stability and accuracy of deep learning-based clinical outcome prediction in lung adenocarcinoma. By reducing technical confounders and enforcing stain consistency, the workflow enhances the robustness, reproducibility, and translational readiness of WSI-derived biomarkers and supports multi-institutional harmonization efforts in computational pathology.