
Evidence on the real-world impact of artificial intelligence (AI) decision support on diagnostic efficiency and workflow outcomes in primary breast cancer pathology remains limited. This retrospective, single-site, paired reader, two-arm study evaluated the impact of AI-assisted digital pathology (DP + AI) versus standard DP on workflow efficiency and diagnostic performance in digitized consecutive breast biopsy cases. Primary endpoints were pathologist review time, lab turnaround time (TAT), and usability (survey-based). Secondary endpoints included diagnostic performance of pathologists across both arms and standalone AI against adjudicated reference standard. Analyses included descriptive statistics, paired t-test, and diagnostic performance metrics with 95% confidence intervals (CIs). The study included 194 digitized breast biopsies. The mean pathologist overall review time was reduced by 43.5% in DP + AI (40.18 [95% CI: 36.20–44.16] s) versus DP (71.12 [95% CI: 64.39–77.84] s) (p < 0.0001). Diagnostic agreement was 97.42% with DP and 99.48% with DP + AI, with 100% sensitivity observed in the DP + AI workflow. The standalone AI showed high diagnostic accuracy, with area under the curve values ≥0.99 across cancer subtypes. Usability survey showed high satisfaction, ease of use, and perceived clinical utility. Baseline TAT analysis of 3298 cases, independent of study cohort, showed longer TAT for cases requiring immunohistochemistry (IHC; 3.73 days) than for those not requiring (1.17 days). Retrospective workflow modelling suggested that AI-guided IHC pre-ordering could potentially reduce TAT by approximately 2.5 days for patients. DP + AI showed improvements in workflow efficiency while maintaining diagnostic performance. These findings suggest that integration of AI into routine breast pathology workflows may support faster diagnosis, optimized laboratory operations, and improved scalability.
This article provides an overview of digital pathology and its evolving role in modern diagnostic practice. We begin by comparing conventional microscopy with digital pathology and outlining the major components of the digital pathology workflow. We then explore whole-slide imaging, its imaging modalities, scanner technologies, image storage, and image management systems. We then examine the role of artificial intelligence (AI) in digital pathology, discussing current AI applications and the spectrum of computational approaches, from conventional feature-engineering methods to deep learning and multimodal models. Finally, we review the clinical applications of digital pathology, summarize its implementation across US medical centers, and discuss the technical, regulatory, and operational challenges that must be addressed for broader clinical adoption.
Background Mitotic activity is a recognized marker of tumor proliferation, yet its reproducible assessment across whole-slide images (WSIs) remains difficult. We developed a transfer-learning framework for lung squamous cell carcinoma (LUSC) and evaluated whether a patch-derived tumor-area mitosis proportion (TAMP) was associated with survival and molecular tumor characteristics. Methods The analysis combined MIDOG 2022 with TCGA-LUSC, including 478 patients and 512 diagnostic WSIs after quality control. Tumor regions and mitotic figures were reviewed by two pathologists with interobserver agreement of Cohen's κ ≥ 0.85. A ResNet-50 model was trained on MIDOG and subsequently adapted to TCGA-LUSC using Focal Loss and stain-aware augmentation. Target-domain performance was evaluated using patient-level stratified 5-fold cross-validation and an additional WSI-level holdout analysis to address potential leakage from the initial patch-level partitioning. TAMP was calculated from patient-level out-of-fold predictions as the proportion of evaluable tumor patches classified as mitosis-positive. Associations with overall survival, immune and genomic features, transcriptomic programs, predicted drug response, and transcription-factor activity were subsequently examined. Multivariable Cox regression adjusted for age, sex, smoking status, tumor stage, and Eastern Cooperative Oncology Group performance status. Results The model achieved an area under the receiver operating characteristic curve (AUC) of 0.975 (95% confidence interval (CI): 0.965–0.980) on the held-out MIDOG test set. In TCGA-LUSC, fine-tuning increased the initial patch-level AUC from 0.82 to 0.92. Patient-level cross-validation yielded a mean AUC of 0.89 (95% CI: 0.86–0.90), and WSI-level holdout validation yielded an AUC of 0.90 (95% CI: 0.84–0.90). High TAMP was associated with poorer overall survival (log-rank p = 0.008; HR = 1.82, 95% CI: 1.18–2.81) and remained associated with mortality after multivariable adjustment (adjusted HR = 1.68, 95% CI: 1.08–2.62; p = 0.022). Higher TAMP was also associated with lower stromal fraction, selected immune-cell differences, greater indel-derived neoantigen load and indel burden, and enrichment of stem-cell proliferation, growth-factor-response, E2F-target, G2M-checkpoint, and MYC-target programs. Additional associations were observed with predicted sensitivity to selected cell-cycle-targeting compounds and with proliferation-related transcription-factor activity. Conclusions TAMP provides a patient-level, image-derived representation of mitosis-positive tumor regions that was associated with survival and several molecular features in TCGA-LUSC. Its association with overall survival persisted after adjustment for major clinical covariates. External multicenter validation, prespecified threshold evaluation, and prospective assessment are required before clinical use.
Aims Muscle biopsy assessment relies on descriptive features that are difficult to compare across samples. We developed MyoPath to quantify these features and examined their usefulness in clinical biopsies and population-based autopsy material. Methods MyoPath combines Cellpose-SAM and conventional image analysis within QuPath to segment myofibers, fat, nuclei, and connective tissue. We analyzed 478 hematoxylin and eosin images from HuashanMuscle (79 snap-frozen biopsies) and GTEx (399 formalin-fixed, paraffin-embedded autopsy samples). Thirty-seven features were summarized by seven pathology indicators. A composite model trained on GTEx healthy and wasting strata was transferred to HuashanMuscle without retraining. Results Segmentation Dice coefficients ranged from 0.87 to 0.95 on annotated cryosections. Nuclear centralization index (NCI) and fiber size coefficient of variation (fiber CV) discriminated myopathy from controls in HuashanMuscle (AUC 0.85 and 0.79) and wasting from healthy donors in GTEx (0.74 and 0.74). Their GTEx associations changed little after age and sex adjustment. The composite achieved AUC 0.788 (95% CI 0.71–0.85) in GTEx and 0.873 (0.78–0.94) in HuashanMuscle. In the latter cohort, mean predicted probability was 0.49 compared with an observed myopathy proportion of 0.79. Conclusions NCI and fiber CV provide interpretable measures of muscle pathology in two distinct tissue settings. MyoPath supports quantitative comparison between groups; clinical severity grading and interpretation of absolute scores require further validation.