Residency · Residency · Pathology
Artificial Intelligence in Pathology: Current Applications and Limitations
Introduction
Artificial intelligence (AI) and machine learning are rapidly entering pathology practice, promising to augment diagnostic accuracy, improve efficiency, and enable quantitative analyses beyond human capability. Understanding the technology, current FDA-cleared applications, validation requirements, and limitations is essential for pathologists navigating this evolving landscape.
Foundational Concepts
Machine Learning Fundamentals
Machine learning (ML) encompasses algorithms that learn patterns from data without explicit programming. Supervised learning trains on labeled data, such as images annotated by pathologists as tumor versus non-tumor. Unsupervised learning identifies patterns in unlabeled data, for example clustering gene expression profiles. Deep learning is a subset of ML that uses multi-layered neural networks, typically convolutional neural networks (CNNs) for image analysis. Data must be split into training, validation, and test sets to prevent overfitting, with the test set evaluating generalizability.
Convolutional Neural Networks in Pathology
CNNs automatically learn hierarchical image features including edges, textures, shapes, and complex patterns. The architecture comprises an input layer, multiple convolutional layers for feature extraction, pooling layers for dimensionality reduction, fully connected layers, and an output layer. Common architectures include ResNet, VGG, Inception, EfficientNet, and U-Net (for segmentation). Transfer learning involves fine-tuning pre-trained models (such as those trained on ImageNet) on pathology images, which reduces training data requirements. Vision transformers, an emerging architecture using self-attention mechanisms, show promise for WSI analysis.
Key Performance Metrics
Sensitivity (recall) is the proportion of true positives correctly identified, while specificity is the proportion of true negatives correctly identified. The area under the ROC curve (AUC) measures overall discriminative ability, where 1.0 is perfect and 0.5 is random. Positive predictive value (PPV) and negative predictive value (NPV) are context-dependent and influenced by prevalence. The F1 score, the harmonic mean of precision and recall, is particularly useful for imbalanced datasets.
FDA-Cleared AI Applications
Current Approved Products
Paige Prostate (Paige AI) was the first FDA-cleared AI product in pathology (2021), detecting areas suspicious for cancer on prostate biopsies and designed as an assistive tool for pathologists. Paige Breast identifies areas of potential cancer on breast biopsies. Proscia DermAI assists in identifying potential skin cancers. Hologic Genius Digital Diagnostics provides AI-assisted cervical cytology screening. PathAI offers algorithms for PD-L1 scoring and HER2 scoring through partnerships with pharmaceutical companies for companion diagnostics. Multiple additional products are in development for IHC quantification, mitosis detection, and tumor grading.
Regulatory Pathway
Most AI pathology tools are cleared as Class II medical devices via 510(k) or De Novo pathways, and must demonstrate substantial equivalence or meet special controls. The Software as a Medical Device (SaMD) framework governs AI algorithms. A predetermined change control plan allows the FDA to manage AI algorithm updates without requiring full resubmission. Post-market surveillance is required to monitor real-world performance.
Clinical Applications
Cancer Detection and Screening
In prostate cancer detection, AI identifies Gleason patterns on needle biopsies, reducing missed cancers and improving grading consistency. For breast cancer metastasis detection, AI screens sentinel lymph node slides and flags suspicious regions for pathologist review. In cervical cytology, AI-assisted screening reduces false negatives and prioritizes slides for review. AI is also being applied to gastric and colorectal cancer for detection of dysplasia, cancer, and microsatellite instability directly from H&E slides.
Quantitative Biomarker Assessment
Automated Ki-67 proliferation index counting is more reproducible than manual estimation. PD-L1 scoring (tumor proportion score, combined positive score) can be standardized across pathologists. Automated HER2 scoring classifies IHC staining intensity and pattern. Tumor-infiltrating lymphocyte (TIL) quantification provides standardized assessment for prognostic and predictive value. Mitotic figure detection assists in identifying and counting mitoses for grading purposes.
Computational Pathology and Predictive Models
Mutation prediction from H&E uses AI models to predict molecular features such as MSI status, BRAF mutation, and TP53 mutation directly from morphology. Outcome prediction models are based on tissue morphology, spatial features, and the tumor microenvironment. Tissue-of-origin prediction applies AI to metastases of unknown primary. Digital spatial profiling integration combines AI morphology analysis with spatial transcriptomics and proteomics.
Validation and Implementation Considerations
Clinical Validation Requirements
Performance must be demonstrated on independent, diverse test datasets not used in training. Potential biases related to tissue processing differences, staining variability, scanner differences, and patient demographics must be addressed. Site-specific validation is essential because algorithms trained at one institution may not perform equally at another due to domain shift. Prospective studies comparing AI-assisted versus standard pathologist workflows are needed. The intended use must be clearly defined, whether as a screening aid, second reader, quantification tool, or autonomous diagnostic system.
Integration into Practice
AI should function as an assistive tool that augments pathologist decision-making rather than replacing it. Workflow integration requires AI results to be presented within the digital pathology viewing platform. The pathologist remains the final decision-maker and is responsible for the diagnosis. The laboratory must validate AI tools under its own conditions per CLIA requirements, and staff must be trained on appropriate use and interpretation of AI outputs.
Limitations and Challenges
Technical Limitations
Generalizability suffers when algorithms are applied to data different from training data, a problem known as domain shift. Staining variability across laboratories with different fixation, processing, and staining protocols degrades performance. Scanner variability introduces color and resolution differences between WSI platforms. Rare entities are poorly represented in training data, leading to poor performance on uncommon tumors or unusual presentations. Explainability remains a challenge, as deep learning models are often "black boxes" whose predictions are difficult to interpret.
Data and Bias Concerns
Training data bias means that if training data lacks diversity in demographics, tissue types, or staining protocols, the model may perform poorly on underrepresented populations. Label quality limits AI performance, since inter-observer variability among annotating pathologists introduces noise into the ground truth. Class imbalance leaves rare diagnoses underrepresented in training data, requiring augmentation or specialized training strategies. Data privacy and governance require de-identification of patient data and compliance with HIPAA and institutional policies.
Practical Barriers
Cost considerations include scanner acquisition, computational infrastructure (GPU servers or cloud computing), and software licensing. Workflow disruption arises because AI requires a transition to fully digital pathology as a prerequisite. Liability questions surround the medicolegal implications when AI contributes to diagnostic decisions. Pathologist acceptance depends on transparent performance reporting and gradual integration to build trust. Algorithm maintenance requires ongoing monitoring for performance degradation from data drift, with periodic retraining.
Clinical Pearls
AI in pathology currently functions as an assistive technology to augment pathologist diagnostic capabilities, not as a replacement for pathologist judgment and responsibility. Site-specific validation is essential because AI algorithms trained at one institution may not generalize to another due to differences in tissue processing, staining, and scanning (domain shift). Quantitative biomarker assessment (Ki-67, PD-L1, HER2) is one of the most immediately impactful applications of AI, offering improved reproducibility over subjective manual scoring. Pathologists should actively engage in AI development and validation to ensure these tools are clinically relevant, rigorously tested, and safely implemented in practice.
References
- Niazi MKK, Parwani AV, Gurcan MN. Digital pathology and artificial intelligence. Lancet Oncol. 2019;20(5):e253-e261.
- Bera K, et al. Artificial intelligence in digital pathology: new tools for diagnosis and precision oncology. Nat Rev Clin Oncol. 2019;16(11):703-715.
- Pantanowitz L, et al. An artificial intelligence algorithm for detecting invasive breast cancer in whole slide images of sentinel lymph node biopsies. Mod Pathol. 2022;35(12):1894-1903.
- FDA. Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices. Available at: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices.