
Background Whole-slide imaging has transformed pathology workflows worldwide, but most implementation studies focus on technical validation or user perceptions, whereas the organizational impact of robotic workflow automation remains insufficiently studied. We evaluated whether robotic end-to-end workflow automation, rather than digitalization alone, could reduce mapped manual handling and estimated technical staffing requirements while supporting positive user evaluations in a high-volume histopathology laboratory. Methods A single-center implementation evaluation combined workflow mapping and institutional operational data with two structured anonymous online questionnaires administered immediately before and 6 months after institute-wide implementation of digital pathology at the Institute of Pathology, Klinikum Stuttgart. Questionnaire data were analyzed descriptively, and free-text responses were grouped thematically. Results The institute processed approximately 230,000 slides annually without dedicated scanning personnel. The process mapping revealed that the robotic workflow involved fewer manual interactions with the slides (about 2 compared to 7 in a sample case involving an immunohistochemistry run); however, these were schematic process figures and not measured workload data. Institutional workforce planning estimated approximately 0.7 full-time-equivalent lower required technical capacity. Twenty-five questionnaires from the pre-implementation period and 21 from the post-implementation period provided usable data; it was not possible to match the respondents across the two time periods. Post-implementation responses were broadly positive. In descriptive, unpaired group-level comparisons, mean ratings for perceived efficiency and productivity were higher after implementation than before implementation. Major challenges included dependence on information technology infrastructure, occasional system instability, and perceived differences in image interpretation compared with conventional microscopy. Conclusions In this single-center implementation, end-to-end robotic workflow automation was compatible with institution-wide digital pathology without dedicated scanning personnel. Process mapping suggested fewer manual slide-touch points, institutional workforce planning estimated lower required technical capacity, and post-implementation user evaluations were broadly positive. These exploratory findings suggest that workflow architecture may contribute to the organizational impact of digital pathology and warrant prospective validation using objective operational and staffing metrics.
Background Prostate cancer is historically considered an immunologically cold tumor. However, the proximity and spatial relevance of T cells to cancer are underexplored. In this study, we utilized artificial intelligence (AI)-based resources to characterize the distribution CD4+ and CD8+ T cells across whole mount radical prostatectomy (RP) slides. Methods Specimens from 105 patients undergoing RP from a single center were reviewed. Hematoxylin and eosin (H&E), anti-CD4 immunohistochemistry (IHC), and anti-CD8 IHC staining from sequential sections were obtained and digitized. Publicly available tools were used to identify and segment cells. A random forest-based classifier was trained to classify positive, negative, and artifactually stained cells. A previously developed H&E cancer detection AI model and DeeperHistReg was used to register tumor borders to IHC slides. Positive cell counts were established for intratumor, tumor border, and extratumoral regions, and thereafter were adjusted for tumor size and correlated with clinical and pathological findings. Results Random forest classifiers achieved F1 scores of 0.88 and 0.72 for CD4+ and CD8+ T cells, respectively, in identification of positive cells. Following registration of H&E to IHC for cancer assessment, there was low correlation between the numbers of intra- and extratumoral CD4+ and CD8+ T cells, suggesting that immune infiltration patters differed per patient. When adjusting for tumor size, we found significant differences in the distribution of CD4+ and CD8+ T cells in cases of lymphovascular infiltration and high prostate-specific antigen values. Conclusion Leveraging AI-based tools, quantitative estimates of T-cell distribution demonstrate the spatial distribution and proximity to cancer may have broad clinicopathological relevance. Given the co-occurrence of many of the clinicopathological variables we studied, it will be important to repeat these studies in larger cohorts to eliminate confounding variables.
Introduction Patient-based real-time quality control (PBRTQC) uses patient results to monitor analytical performance, but performance depends strongly on the disturbance type and monitoring parameters. This study compared a tuned multianalyte moving-average PBRTQC method with logistic regression, Random Forest, and XGBoost for detection of simulated clinical chemistry disturbances and evaluated transportability to independently sourced National Health and Nutrition Examination Survey (NHANES) chemistry profiles. Materials and methods Independent synthetic training and test cohorts of 10,000 observations each were generated for glucose, sodium, potassium, and creatinine. Error prevalence was 10% in training and 5% in testing. Error-positive observations were assigned to systematic shift, progressive drift, hemolysis-associated, or delayed-processing scenarios. Machine-learning (ML) hyperparameters and classification thresholds were selected using training data only; thresholds were derived from 5-fold out-of-fold predictions. PBRTQC window size and control limit were tuned on the training cohort. Transportability was evaluated using 400 complete NHANES August 2021–August 2023 participant profiles. Each profile was represented under normal, shift, drift, hemolysis, and mixed scenarios. Because the 2000 scenario observations arose from 400 participants, uncertainty for the overall NHANES analysis was estimated by 1000-replicate participant-level cluster bootstrap. Results In the independent synthetic test cohort, ROC-AUC was 0.673 for logistic regression, 0.786 for Random Forest, 0.797 for XGBoost, and 0.536 for tuned PBRTQC. Performance varied markedly by disturbance type; for example, hemolysis-associated AUCs were 0.985, 0.978, 0.975, and 0.642, respectively, whereas drift AUCs were 0.643, 0.665, 0.689, and 0.518. In the NHANES-based transportability analysis, overall AUC was 0.886 (95% cluster-bootstrap CI 0.873–0.900) for logistic regression, 0.761 (0.741–0.779) for Random Forest, 0.795 (0.777–0.812) for XGBoost, and 0.807 (0.793–0.823) for PBRTQC. PBRTQC had high sensitivity (0.977) but low specificity (0.079) in the constructed NHANES scenario set. Conclusions Detection performance was disturbance-dependent, and no single method was uniformly superior across scenarios. ML models showed stronger internal discrimination for several sample-level perturbations, whereas PBRTQC performed strongly for some NHANES-derived perturbation scenarios but generated many false-positive classifications at the selected operating point. The NHANES analysis should be interpreted as a transportability evaluation using real patient chemistry distributions with computationally imposed disturbances, not as validation against documented real-world quality-control failures.