Pathology and Laboratory Medicine

Pathology AI applies computational methods to digitized slides, laboratory data, and workflow operations. Its clinical value depends on the intended use, authorized workflow, scanner and stain compatibility, case mix, pathologist interaction, and prospective monitoring, not image-classification performance alone. A pathology model is one component of a specimen-to-report system; the pathologist still integrates morphology, stains, molecular results, specimen quality, and clinical context.

Learning Objectives

After reading this chapter, you will be able to:

  • Evaluate digital pathology and histopathology systems, including authorized prostate-detection and research second-read workflows
  • Critically assess AI applications in cytopathology (Pap smears, thyroid FNA), gynecologic histopathology (endometrial, ovarian, cervical biopsy), and hematopathology (peripheral blood smears, bone marrow analysis)
  • Understand the current state of macroscopic gross pathology AI and why it lags digital histopathology
  • Understand AI tools for clinical laboratory workflow optimization, quality control, and critical value detection
  • Analyze the unique regulatory landscape for laboratory AI, including FDA clearance pathways and CLIA implications
  • Recognize failure modes specific to pathology AI, including scanner variability, staining heterogeneity, and rare diagnosis challenges
  • Navigate medico-legal implications when AI assists (or misses) pathologic diagnoses
  • Apply evidence-based frameworks for evaluating pathology AI before clinical deployment

The Clinical Context:

Pathology may be the medical specialty most transformed by AI in the next decade. The transition from glass slides to digital pathology (whole-slide imaging) has created millions of digitized specimens, rich training data for computer vision algorithms. Pathology is fundamentally pattern recognition: identifying cancer cells among normal tissue, grading tumor differentiation, quantifying biomarker expression. These are tasks where AI excels.

Unlike radiology (where imaging modalities vary) or clinical medicine (where symptoms are subjective), pathology has: - Standardized specimens: H&E staining protocols, immunohistochemistry methods - Archived training data: Decades of diagnosed cases with known outcomes - Clear ground truth: Biopsy diagnoses confirmed by clinical outcome - Quantifiable features: Cell count, nuclear-to-cytoplasmic ratio, mitotic figures

Result: Pathology AI includes authorized devices, retrospective validation, reader studies, prospective workflow evaluation, and research models. These evidence categories are not interchangeable.

But challenges remain: Scanner variability, staining heterogeneity, rare diagnoses with limited training data, and medico-legal questions about liability when algorithms miss critical diagnoses.

Key Applications:

  • Prostate cancer detection: Paige Prostate Detect is De Novo authorized for adjunctive suspicious-focus detection, not Gleason scoring. Ibex second-read performance belongs to a separate system.
  • Breast cancer HER2 scoring: PathAI describes AIM-HER2 as research use only. Platform authorization must not be transferred to the scoring algorithm.
  • Lymph-node metastasis detection: LYNA reader-assistance evidence showed improved sensitivity and review time in one research workflow; it was not an autonomous clinical deployment.
  • Foundation-model aggregators: nnMIL is a slide-level MIL connector across about 40,000 WSIs and 35 tasks, not a clearance or outcomes claim (Luo et al., 2026)
  • Cervical cytology: Pooled estimates combine heterogeneous tasks and designs; product performance requires product-specific evidence.
  • Gynecologic histopathology: Retrospective prognostic associations do not establish treatment selection or clinical actionability.
  • Macroscopic gross pathology: A lesion-localization benchmark does not establish improved sampling or diagnosis.
  • Workflow optimization and case triage: AI flags high-risk cases for expedited pathologist review, reduces turnaround time
  • Peripheral blood smear analysis: Automated differential counts and blast detection, variable performance across analyzers
  • Bone marrow biopsy analysis: Research evidence is heterogeneous; regulatory status must be checked for the exact product and intended use
  • Rare tumor diagnosis: AI trained on common cancers performs poorly on rare subtypes (sarcomas, rare lymphomas)
  • Autonomous diagnosis without pathologist review: Broad clinical evidence is not established; the permitted role and required review are product-specific
  • “One-size-fits-all” algorithms across scanner types: Compatibility across scanners and image pipelines requires direct evidence

What Actually Works:

  1. Paige Prostate Detect: Product-specific De Novo authorization for adjunctive suspicious-focus detection
  2. Ibex prostate second read: External and routine-workflow evidence that must not be attributed to Paige
  3. LYNA breast metastasis reader study: Sensitivity increased from 83% to 99%, while mean review time decreased from 116 to 61 seconds in the evaluated workflow (Steiner et al., 2018)
  4. Prospective prostate workflow: CONFIDENT-P reduced immunohistochemistry use but did not significantly reduce assessment time or establish better diagnostic outcomes (Flach et al., 2025)

What Does Not Work:

  1. Scanner-agnostic algorithms: AI trained on Aperio scanners often fails on Leica or Hamamatsu; each vendor requires separate validation
  2. Algorithms for rare diagnoses: Angiosarcoma, desmoplastic melanoma, rare lymphoma subtypes lack training data; AI confidently misdiagnoses
  3. Stain-independent analysis: H&E algorithms fail when staining intensity varies; requires standardized protocols
  4. Unbounded autonomy: An authorization or reader study cannot be generalized beyond its labeled review model

Critical Insights:

Pathology AI is task-specific: Detection, grading, quantification, triage, prognosis, and quality control have different intended uses and evidence requirements.

Scanner standardization is the hidden challenge: Algorithms validated on one scanner brand often perform poorly on others; this limits portability

Quantitative immunohistochemistry is AI’s sweet spot: HER2 scoring, PD-L1 tumor proportion score, Ki-67 proliferation index. AI reduces inter-observer variability where pathologists already struggle

Rare and unfamiliar diagnoses require separate testing: Common-condition performance does not establish safe behavior on rare tumors, mimics, or artifacts. A 21-pathologist FHdRCC reader study showed a median accuracy gain with AI that was significant in GU specialists and not in other groups (Park et al., 2026).

Gastric-biopsy multi-class WSI models remain reader-study evidence: GBAIM reported high HGD+ sensitivity and specificity on Chinese external and prospective slides; that is not a staging or autonomous-signout claim (Chen et al., 2026).

Reusable patch embeddings still need a validated aggregator: A foundation-model feature extractor is not a validated clinical biomarker until the aggregator, external cohorts, and intended-use endpoint are specified.

Workflow and accuracy answer different questions: Turnaround time, ancillary testing, diagnostic agreement, and patient outcomes should be measured separately.

Medico-legal landscape is evolving: If AI misses a cancer, is the pathologist liable for not catching the algorithm’s error? If pathologist disagrees with AI and turns out wrong, is that negligence?

Clinical Bottom Line:

Pathology AI has clinically relevant authorized and research applications. Adopt an exact product and workflow only when the evidence, authorization, local validation, and monitoring fit the intended task.

For common diagnoses with abundant training data, selected systems can support detection or quantification, but class-wide superiority to pathologists is not established.

For rare diagnoses, complex cases, and situations where clinical context matters, the correct response may be deferral, ancillary testing, consultation, or complete independent review. A confidence score is not a calibrated probability of diagnosis unless the study establishes that interpretation.

Demand evidence: - Ask vendors for FDA clearance documentation (not just submission) - Request scanner-specific validation data (your exact scanner model) - Prespecify a risk-based local validation sample and acceptance criteria appropriate to the intended use - Monitor performance stratified by diagnosis, specimen type, and staining protocol

Medico-Legal Considerations:

  • Responsibility is context-specific: Duties can involve the pathologist, laboratory, institution, vendor, and others depending on jurisdiction, labeling, policy, contract, and facts
  • Standard of care is jurisdiction- and fact-specific: Do not assume that use or nonuse of an available tool is automatically negligence
  • Documentation requirements: Note AI assistance in pathology report; document when you disagree with algorithm
  • FDA authorization is not a superiority claim: The pathway, intended use, controls, and evidence must be checked in the exact record
  • Informed consent unclear: Should patients be told their pathology was AI-assisted? Current practice: No, but this may change
  • Liability for false negatives: Allocation depends on the facts and jurisdiction; no categorical answer follows from who signed the report
  • Liability for false positives: If AI flags benign tissue as cancer and pathologist disagrees but orders unnecessary treatment, is relying on clinical judgment adequate defense?

Essential Reading:

  • Pantanowitz L et al. (2020). “An artificial intelligence algorithm for prostate cancer diagnosis in whole slide images of core needle biopsies: a blinded clinical validation and deployment study.” Lancet Digital Health 2:e407-e416. [Ibex prostate algorithm validation and deployment]

  • Steiner DF et al. (2018). “Impact of Deep Learning Assistance on the Histopathologic Review of Lymph Nodes for Metastatic Breast Cancer.” American Journal of Surgical Pathology 42:1636-1646. (Steiner et al., 2018) [LYNA breast metastasis detection; AI-assisted pathologists achieved 91% sensitivity vs. 83% unassisted for micrometastases]

  • Bao H et al. (2020). “Artificial intelligence-assisted cytology for detection of cervical intraepithelial neoplasia or invasive cancer: A multicenter, clinical-based, observational study.” Gynecologic Oncology 159:171-178. (Bao et al., 2020) [Cervical cytology AI validation study]

  • Hanna MG et al. (2020). “Validation of a digital pathology system including remote review during the COVID-19 pandemic.” Modern Pathology 33:2115-2127. (Hanna et al., 2020) [Digital pathology workflow validation; 100% major diagnostic equivalency between digital and glass]

  • Lin E et al. (2023). “Digital pathology and artificial intelligence as the next chapter in diagnostic hematopathology.” Seminars in Diagnostic Pathology 40:88-94. (Lin et al., 2023) [Hematopathology AI review]

  • Soleymanjahi S et al. (2024). “Artificial Intelligence–Assisted Colonoscopy for Polyp Detection: A Systematic Review and Meta-analysis.” Annals of Internal Medicine 177:1652-1663. (Soleymanjahi et al., 2024) [AI-assisted colonoscopy increases polyp detection; impacts upstream specimen volume for pathologists]

Introduction: Clinical Scope and Evidence

Pathology is pattern recognition. A pathologist examines a biopsy slide, identifies malignant cells among normal tissue, grades tumor differentiation, counts mitotic figures, and integrates these visual features with clinical context to render a diagnosis.

This is exactly what computer vision AI does: detect patterns in images.

Unlike clinical medicine (where symptoms are subjective and examination findings nuanced), pathology offers:

1. Digitized specimens: Whole-slide imaging scanners convert glass slides to gigapixel digital images

2. Abundant training data: Decades of archived specimens with confirmed diagnoses

3. Standardized protocols: H&E staining, immunohistochemistry methods largely consistent across labs

4. Quantifiable features: Cell count, nuclear size, mitotic index, staining intensity can be measured objectively

5. Clear ground truth: Biopsy diagnoses confirmed by clinical outcome, surgical pathology, molecular testing

The result: Pathology AI has progressed from retrospective image classification to authorized detection aids, reader studies, and some prospective workflow evaluation. The strength of evidence still varies substantially by product and task.

But pathology AI faces unique challenges:

  • Scanner variability: Algorithms trained on Aperio scanners may fail on Leica or Hamamatsu
  • Staining heterogeneity: H&E staining intensity varies across labs, confusing algorithms
  • Rare diagnoses: Limited training data for uncommon tumors, leading to confident misdiagnoses
  • Clinical context: AI sees pixels, not patient history (the chest wall mass in a 70-year-old smoker is likely different from identical-appearing mass in 25-year-old)

The emerging paradigm: foundation models

The field is shifting from task-specific algorithms toward foundation models trained on large pathology datasets. UNI, CONCH, GigaPath, mSTAR, and Atlas use self-supervised or multimodal learning and can be adapted for cancer classification, subtyping, outcome prediction, and biomarker research (Xu et al., 2025). Broader task coverage by design does not establish clinical generalization or superiority to a specialized model.

Real-world deployment signal: Campanella and colleagues fine-tuned an open-source pathology foundation model to predict EGFR mutation status from H&E lung adenocarcinoma slides in an international dataset of 8,461 slides. The model achieved mean AUC 0.847 internally and 0.870 externally, then AUC 0.890 in a prospective silent trial on primary samples. The AI-assisted workflow could reduce rapid molecular tests by up to 43% while maintaining current clinical-standard performance (Campanella et al., 2025).

Clinical interpretation: This is a stronger translation signal than benchmark-only foundation-model papers because it includes external validation and prospective silent deployment. It still does not replace molecular testing; it triages tissue-preserving testing strategy and should be monitored by site, scanner, stain, tumor cellularity, and mutation prevalence.

A 2026 Nature Biomedical Engineering study introduced nnMIL, a multiple-instance learning aggregator that connects pathology foundation-model patch features to slide-level prediction through random patch and feature sampling and sliding-window ensemble inference (Luo et al., 2026). Across about 40,000 WSIs, 35 tasks, and 4 foundation models, the authors report consistent gains over prior MIL methods for diagnosis, subtyping, biomarker detection, and pan-cancer prognosis. Treat nnMIL as an aggregator and benchmark result, not an outcomes or clearance claim.

PathPrism encodes colorectal-cancer whole-slide images into 628 interpretable spatial features (tissue fractions, spatial entropy, and graphs) and, in a retrospective 7,000-patient study across 11 cohorts, associated those features with disease-specific survival on external MCO, CR07, and TCGA slides and with MSI, BRAF, and TP53 status (Liang et al., 2026). Language models organized the spatial signatures into expert-reviewed hypotheses, and VirtualWSI applied in silico semantic perturbation to interrogate the model, not to establish biological causality. Cox interaction analyses that stratified adjuvant-chemotherapy associations in stage II/III disease remain hypothesis-generating retrospective findings, not a predictive treatment-benefit test.

Prospective workflow evidence: The CONFIDENT-P study prospectively evaluated an AI-supported prostate-biopsy workflow in 82 patients and 237 slides using an alternating implementation design. AI support reduced immunohistochemistry use, but the study did not show a significant reduction in assessment time and did not establish improved diagnostic outcomes (Flach et al., 2025). Reduced ancillary testing is a workflow result, not proof of better diagnosis or patient outcome.

Part 1: Histopathology AI. Prostate and Breast Cancer

Prostate Cancer Detection and Second Read

Prostate cancer diagnosis from needle core biopsies is one of the most common pathology specimens. Gleason grading (scoring architectural patterns) predicts prognosis and guides treatment. But Gleason grading has notorious inter-observer variability. Different pathologists assign different scores to the same slide.

The AI solution: Computer vision algorithms trained on tens of thousands of digitized prostate biopsies learn to detect cancer foci and assign Gleason patterns.

Paige Prostate Detect (De Novo DEN200080)

FDA authorized Paige Prostate through the De Novo pathway on September 21, 2021. It is an adjunctive aid for detecting suspicious foci after a pathologist’s initial review of digitized H&E-stained, formalin-fixed paraffin-embedded prostate needle-biopsy slides. The authorized configuration uses the Philips IntelliSite Ultra Fast Scanner and Paige FullFocus viewing software, and the output is not intended as the primary diagnosis (FDA DEN200080).

The authorized use is suspicious-focus detection, not Gleason scoring, grading, treatment selection, or autonomous diagnosis. Performance figures from another prostate algorithm cannot be assigned to Paige.

Clinical validation of a second-read prostate algorithm: The most fully reported blinded validation and routine deployment described here covers an Ibex system rather than Paige Prostate. Pantanowitz et al. (2020) validated it on an external set of 100 consecutive cases (1,627 slides), then ran it as a second-read system in routine practice (Pantanowitz et al., 2020): - AUC 0.991 (95% CI 0.979-1.00) for cancer detection in external validation, and 0.997 (0.995-0.998) in internal testing - AUC 0.957 for perineural invasion; algorithm and pathologist cancer percentage agreed at r=0.882 - In routine use across 11,429 slides from 941 cases, the system generated 560 cancer alerts and 90 Gleason 7-10 alerts, and detected a cancer that had been missed on initial review

FDA later cleared Ibex’s Galen Second Read under K241232 on January 24, 2025. The exact FDA record should govern claims for the cleared version and workflow (FDA K241232). The Pantanowitz evidence and Galen authorization remain separate from Paige’s evidence.

How pathologists use it: 1. Prostate needle biopsy digitally scanned (whole-slide imaging) 2. The authorized detection aid analyzes the slide and identifies a suspicious location when its criteria are met 3. Pathologist reviews flagged areas + entire slide 4. Pathologist renders final diagnosis (algorithm is advisory, not determinative)

Current deployment: Commercial ownership, laboratory counts, and monthly volume can change and should be verified from current primary records. A corporate acquisition or deployment claim does not expand the FDA-authorized indication.

Why this works: - Common diagnosis: Prostate cancer is most common male cancer; abundant training data - Standardized specimen: Needle core biopsies processed similarly across labs - Clear patterns: Gleason architecture (glands, cribriform, solid sheets) are well-defined visual features - Addresses a defined problem: Suspicious-focus detection can be evaluated separately from grading and prognosis

Limitations: - Scanner-specific: Validated on Philips IntelliSite Ultra Fast scanners; performance on other scanners requires separate validation - Task boundary: The authorized Paige detection aid does not assign Gleason grade - Mimics: Atrophy, adenosis, atypical adenomatous hyperplasia can fool algorithm (and pathologists)

Breast Cancer Lymph Node Metastasis Detection

Sentinel lymph node biopsy in breast cancer requires pathologists to scan entire lymph nodes for micrometastases (sometimes <0.2mm clusters of cells). Tedious, error-prone work.

Google/Verily LYNA (Lymph Node Assistant)

Published: Steiner et al. (2018) in AJSP (Steiner et al., 2018)

Training data: 399 whole-slide images of sentinel lymph nodes

Performance: - Sensitivity for micrometastases (≥0.2mm): 99% - False positive rate: 9% (acceptable for triage/flagging tool) - Halves pathologist reading time: Pathologists using LYNA reviewed slides 50% faster with equivalent accuracy

Clinical validation: Six pathologists reviewed 130 lymph node slides without LYNA, then with LYNA: - Without LYNA: 83% sensitivity for micrometastases (17% false negative rate) - With LYNA: 99% sensitivity (1% false negative rate) - Time per slide: 61 seconds (LYNA-assisted) vs. 116 seconds (unaided)

Why this works: - Clear visual target: Metastatic breast cancer cells look different from lymphocytes - Tedious human task: Scanning entire lymph node for rare micrometastases - High clinical stakes: Missing micrometastasis changes staging and treatment

Why this did not deploy widely: - No authorization record cited here: The LYNA publication is research evidence, not an FDA product record - Scanner specificity: Algorithm performance depends on specific scanning protocol - Workflow integration challenges: Many labs have not adopted whole-slide imaging yet

The lesson: Reader-study performance does not establish a generalizable clinical workflow. Authorization, scanner compatibility, external validation, vendor support, and laboratory integration remain separate requirements.

Breast Cancer HER2 Scoring

HER2 immunohistochemistry scoring uses ordinal categories that inform confirmatory testing and treatment eligibility. The applicable interpretation and treatment pathway depend on the tumor setting, assay result, current guideline, and drug labeling. The 2023 ASCO/CAP update retained the existing scoring system and emphasized reporting the semi-quantitative result; a 0 or 1+ result should not be converted into a universal statement that no HER2-directed treatment is indicated (CAP and ASCO HER2 Testing Guideline Update).

The problem: Inter-observer variability. Different pathologists score the same slide differently, especially for borderline 2+/3+ cases.

The AI solution: Quantitative image analysis measures membrane staining intensity objectively.

PathAI AIM-HER2 (Developer-Described Research Use Only):

PathAI launched AIM-HER2 Breast Cancer in July 2023 as an AI-powered HER2 scoring algorithm. Developed using 157,000 tissue annotations and consensus scores from over 65 expert breast pathologists on 4,000+ slides.

Developer-reported research characteristics (PathAI, 2023): - Concordance with expert breast pathologists: Improved inter-rater agreement, particularly at 0/1+ and 1+/2+ cutoffs - Reduces equivocal scoring: Better distinction between borderline cases - Standardizes scoring: Decreases inter-pathologist variability

Current status: AIM-HER2 is described by the developer as research use only and not for diagnostic procedures. FDA cleared AISight Dx under K243391 on June 26, 2025 as software for viewing and managing digital images of scanned FFPE surgical-pathology slides; the record does not authorize AIM-HER2 as a diagnostic scoring algorithm (FDA K243391). In March 2026, PathAI announced Breakthrough Device Designation for PathAssist Derm. FDA generally does not publicly disclose designation before marketing authorization unless the sponsor does, and designation is not authorization (FDA Breakthrough Devices Program; PathAI announcement).

Potential research and workflow questions: - Whether quantitative assistance improves reproducibility at clinically important boundaries - Whether any change in confirmatory testing is guideline-concordant and prospectively measured - Whether performance remains stable across stains, scanners, laboratories, tumor settings, and evolving treatment criteria

AI-Assisted Colonoscopy and Pathology Workflow Impact

While not strictly pathology AI, AI-assisted colonoscopy systems that improve polyp detection directly impact pathology specimen volume.

Evidence from meta-analysis: A 2024 systematic review and meta-analysis of 44 RCTs (36,201 colonoscopies) found that computer-aided detection (CADe) systems significantly increase adenoma detection (Soleymanjahi et al., 2024):

  • Adenoma detection rate (ADR): 44.7% with CADe vs. 36.7% without
  • Advanced colorectal neoplasia (ACN) detection: 12.7% vs. 11.5%
  • Trade-off: Resection of approximately 2 additional nonneoplastic polyps per 10 colonoscopies

Implications for pathologists: - Increased specimen volume: Higher polyp detection translates to more polyp specimens for histopathologic examination - More nonneoplastic polyps: The increased detection includes both neoplastic and nonneoplastic polyps, increasing routine workload - Upstream AI affecting downstream pathology: This represents an important category where AI in one specialty creates workflow changes in pathology

Current status: FDA has authorized specific colonoscopy CADe systems, but those device records and endoscopy outcomes should not be transferred to pathology AI. Their relevance here is the downstream effect on specimen volume.

Gastric biopsy whole-slide multi-class AI

Chen et al. trained GBAIM (HRNet18 with multiple-instance learning) on 20,711 WSIs from 17,086 patients across six centers for six-class NAG / AG / BP / LGD / HGD / GC classification, then prospectively tested 3,698 WSIs from 2,965 patients (Chen et al., 2026). Pooled-external HGD+ sensitivity / specificity was 95.9% (93.7–97.4) / 95.1% (94.7–95.5), with all-external PPV 0.468 and micro-average AUC 0.929; the prospective GZ-1-PR cohort reached HGD+ 95.7% / 99.3%. In a 300-slide, 6-week-washout reader study, assistance raised accuracy for all nine pathologists by 1.7–39.0% and cut time 29.4–50.5%. This is a biopsy classification and reader-study result, not a staging or deployment claim.

Part 1b: Macroscopic Gross Pathology AI

Gross pathology (macroscopic examination of surgical specimens before tissue sectioning) is where pathologists assess specimen orientation, margin involvement, tumor location, and which areas to sample for histology. This step is conventionally underdocumented and highly operator-dependent. AI for gross pathology is an emerging frontier, substantially less mature than digital histopathology but increasingly recognized as a gap worth closing.

What Macroscopic AI Can Do

Surgical specimen lesion localization:

Object-detection and semantic-segmentation models trained on macroscopic photographs of resected gastric cancer specimens achieved 95.9% mean average precision for the reported lesion-localization task (Yang et al., 2021). The clinically relevant question is whether such output improves specimen sampling or diagnostic outcomes. The benchmark did not test that clinical utility claim.

Multiscale integration: gross and microscopic features together:

Modern computational pathology models now extract features at both macroscopic (tissue architecture, gross tumor appearance) and microscopic (cellular) scales within a unified framework. The NCI’s 2025 workshop on digital pathology imaging identified multiscale analysis combining macroscopic and microscopic features as a priority research direction, noting that point-of-care macroscopic AI could enable triage and preliminary assessment before subsequent histopathologic validation (NCI Digital Pathology Workshop, J Pathol Inform, 2025).

Why Macroscopic AI Lags Histopathology AI by Years

Two structural challenges explain the gap:

  1. Image acquisition varies: Gross specimen photography differs by lighting, angle, background, magnification, equipment, and operator. Generalizability requires documented acquisition standards and multicenter testing. Single-site performance should not be converted into a claim that a model will fail everywhere else or work across sites.

  2. Training data was never archived: Decades of glass slides are now being digitized for histopathology AI. Macroscopic photographs were not systematically archived in most departments. The training data problem for gross pathology is far more severe than for microscopy.

Clinical implication: Macroscopic assessment remains a research area. Claims of broad deployment readiness require standardized image acquisition, external validation, and prospective evidence that the workflow improves sampling or diagnostic outcomes. A fixed prediction about when the field will mature is not evidence.

Part 2: Cytopathology AI

Cervical Cytology (Pap Smear) Screening

Pap smear screening for cervical dysplasia has dramatically reduced cervical cancer incidence. But cytotechnologists manually review hundreds of slides per day looking for rare dysplastic cells, tedious work prone to false negatives.

AI applications: - Automated screening: Flag slides with suspected abnormalities for cytotechnologist review - Quality control: Re-review slides initially called negative to catch false negatives - Workflow optimization: Prioritize worklist based on AI-assessed abnormality likelihood

Historical computer-assisted systems: - BD FocalPoint and Hologic ThinPrep imaging illustrate conventional computer-assisted cytology workflows that predated current deep-learning systems. - Their regulatory and clinical records should be verified from the exact FDA PMA or supplement before making a current product claim. - Historical rule-based image analysis is not evidence for modern machine-learning cytology systems.

Performance: Liu et al. (2024) systematic review and meta-analysis of 77 studies on cervical cytology AI (Liu et al., 2024): - Pooled sensitivity for AI-assisted Pap smears: 95% (95% CI 91-98%) - Specificity: 94% (95% CI 89-97%) - Pooled accuracy: 94% (95% CI 92-96%); in colposcopy, AI was more accurate than experienced colposcopists (OR 1.75; 95% CI 1.33-2.31)

How cytotechnologists use it: - AI pre-screens slides, ranks by abnormality likelihood - Cytotechnologist reviews high-risk slides first - AI flags slides initially called negative for second review - Cytotechnologist renders final diagnosis

Potential clinical and operational effects to measure: - Sensitivity for the prespecified lesion threshold - Worklist prioritization and cytotechnologist review time - False-negative review, false-positive workload, colposcopy referral, and downstream testing

Limitations: - Pooled specificity was 94%, not 12%–15%; individual-study false-positive burden still depends on prevalence, threshold, preparation, and workflow - Does not detect all lesions: Glandular abnormalities (adenocarcinoma in situ) often missed - Preparation-dependent: Algorithms validated on ThinPrep may not work on SurePath

Gynecologic Histopathology AI

Gynecologic cancers have been relatively overlooked in digital pathology compared to prostate and breast cancer, despite representing the fourth most common cancer group in women (Joshua et al., Cancers, 2025). AI is now entering the full spectrum: cervical biopsy CIN grading, endometrial subtyping, and ovarian prognosis prediction from H&E whole-slide images.

Cervical biopsy CIN grading on whole-slide images:

Moving beyond cytology smears, research applies deep learning to cervical biopsy grading on H&E whole-slide images. The cited Liu review pooled heterogeneous cervical AI tasks and should not be used as a biopsy-product performance record or as proof that one architecture is clinically superior (Liu et al., 2024).

Endometrial cancer: AI identifies a high-risk subtype invisible to conventional pathology:

Using deep learning on H&E histopathology images, Darbandsari et al. identified a novel subset of NSMP (no specific molecular profile) endometrial cancers, termed “p53abn-like NSMP,” with markedly inferior progression-free and disease-specific survival. This subtype is undetectable by traditional morphological assessment or standard molecular classifiers (Darbandsari et al., Nat Commun, 2024):

  • Discovery cohort: 368 patients; independently validated in two external cohorts (290 and 614 cases)
  • The subgroup’s prognostic association was not a prospective treatment-selection test
  • The study did not establish that patients were undertreated or that model-guided management improves outcomes

Ovarian cancer: digital pathology index predicts prognosis and treatment response:

Yang et al. developed the Ovarian Cancer Digital Pathology Index (OCDPI), a graph-based deep learning model that predicts prognosis and adjuvant therapy response from H&E-stained WSI alone, without molecular testing (Yang et al., Eur J Cancer, 2024):

  • Validated in TCGA-OV discovery cohort and two external cohorts (PLCO and Harbin Medical University Cancer Hospital)
  • OCDPI stratified overall survival: HR 1.92 (95% CI 1.38-2.66, p<0.001) and HR 2.80 (95% CI 1.40-5.57, p=0.002) in external cohorts
  • Treatment-response analyses were retrospective associations and do not establish a predictive biomarker for selecting adjuvant therapy
  • OCDPI added independent prognostic value above standard clinicopathological factors

Gynecologic pathology AI is progressing from detection tasks toward retrospective prognostic and biomarker research. Prognostic separation is not clinical actionability. A model-defined subgroup becomes actionable only after analytical validation, external replication, and evidence that using it improves a clinical decision or outcome.

Part 3: Hematopathology AI

Peripheral Blood Smear Analysis

Complete blood count (CBC) analyzers provide automated cell counts, but peripheral blood smear review by humans identifies: - Blast cells (leukemia diagnosis) - Abnormal red cell morphology (hemolytic anemia) - Parasites (malaria, babesiosis) - Left shift (bands, myelocytes indicating infection)

AI applications: - Automated differential counts: Classify neutrophils, lymphocytes, monocytes, eosinophils, basophils - Blast detection: Flag suspected leukemia cases - Malaria detection: Quantify parasitemia in endemic areas

Performance: Performance varies by analyzer, cell type, specimen preparation, disease spectrum, prevalence, and reference standard. A review found promising results across automated differentials, abnormal-cell flagging, and parasite detection, but did not support one class-wide performance range or a universal superiority claim (Lin et al., 2023).

Current use: - Research stage for most applications - Manual review requirements depend on the exact analyzer, labeling, result type, laboratory procedure, and abnormality - Malaria systems require validation against the intended microscopy, specimen, parasite density, and clinical workflow

Limitations: - Morphologic subtlety: Distinguishing reactive lymphocytes from lymphoma cells requires expertise AI lacks - Rare cells: Abnormal cells (blasts, atypical lymphocytes) are infrequent; limited training data - Preparation artifacts: Smear quality (cell distribution, staining) affects algorithm performance

Bone Marrow Biopsy Analysis

Bone marrow analysis requires: - Cellularity assessment: Hypocellular, normocellular, hypercellular - Blast count: Blast percentage contributes to classification within current diagnostic criteria; genetic and disease-specific criteria can alter how thresholds are applied - Cellular morphology: Dysplasia, maturation abnormalities, infiltrative process

AI research: Multiple academic studies evaluate automated blast counting, cellularity assessment, and megakaryocyte quantification. Performance must be tied to the specimen, disease spectrum, reference annotation, and classification task rather than a universal correlation range.

Why this is not clinically deployed yet: - Heterogeneity: Bone marrow has far more cell types than peripheral blood - Context matters: Blast count interpretation requires clinical history (prior chemo, transplant status) - Rare diagnoses: Hairy cell leukemia, systemic mastocytosis, metastatic disease have limited training data - Regulatory and clinical status should be verified for the exact product rather than inferred from the research literature

Part 4: Clinical Laboratory AI and Workflow Optimization

Quality Control and Critical Value Detection

Clinical chemistry and hematology labs already use rules-based algorithms extensively: - Delta checks: Flag results that changed dramatically from prior (possible specimen mix-up or critical illness) - Critical value alerts: Automatic notification for laboratory-defined critical thresholds - Quality control monitoring: Levey-Jennings charts, Westgard rules

AI enhancements: Machine learning improves these by learning lab-specific patterns: - Predict when QC will fail (proactive instrument maintenance) - Reduce false positive delta check alerts (by learning patient-specific patterns) - Identify systematic errors (reagent lot problems, instrument drift)

Current use: Research and proprietary implementations in large reference labs. Not yet widely adopted in community hospital labs.

Microbiology AI

Applications: - Automated colony counting: Computer vision counts bacterial colonies on agar plates - Species identification from imaging: Distinguish E. coli from Klebsiella based on colony morphology - Antibiotic resistance prediction: Predict susceptibility from genomic data (whole-genome sequencing) - Blood culture positivity prediction: Predict which blood cultures will grow bacteria (to prioritize processing)

Performance: Schinkel et al. (2023) ML model for blood culture stewardship: - Predicted blood culture outcomes with AUC 0.75–0.81 in prior emergency department validations using laboratory results and vital signs, and held a stable average AUC of 0.78 across 12 months of real-time prospective monitoring - Identified lower predicted-risk patients for stewardship research; prediction alone did not establish that withholding cultures was safe

Current use: - Automated colony counting: Product- and laboratory-workflow specific - Genomic resistance prediction: Research stage, expensive (WGS not routine) - Blood culture prediction: Research stage

Part 5: Scanner and Image-Pipeline Compatibility

Algorithm performance can change across scanners, software versions, stains, and image-processing pipelines. Compatibility must be demonstrated for the deployed configuration.

A change from one scanner or image pipeline to another can alter inputs even when the underlying tissue is identical. Important mechanisms include:

Technical reasons: - Color space differences: Scanners capture RGB values differently - Compression artifacts: JPEG compression algorithms vary - Focus algorithms: Autofocus creates slightly different focal planes - Illumination: LED vs. halogen light sources, color temperature differences

Clinical consequences: - Labs must validate algorithms on their specific scanner before deployment - Switching scanners or image-processing versions may require compatibility assessment or revalidation for each deployed algorithm - Scanner-independent claims require direct multiscanner validation for the intended workflow

Hypothetical compatibility failure: A laboratory deploys prostate cancer detection AI on one validated scanner configuration and later changes scanners or image-processing software. After the change, staff observe more discordant alerts and a clinically important false negative. The event, performance change, and causal mechanism are fictional teaching details, not a published case report.

They had to: 1. Halt clinical use 2. Confirm whether the new configuration is within the authorization and validation record 3. Conduct risk-based acceptance testing with representative local cases and prespecified criteria 4. Review discordant cases and document the decision to resume, restrict, or discontinue use

The timeline and cost depend on the change, risk, specimen mix, case availability, vendor support, and laboratory quality system. No fixed six-month delay or $50,000 remediation cost should be presented as a real event.

The lesson: Ask vendors explicitly: “Was this algorithm validated on our exact scanner model and software version?”

Part 6: Rare and Out-of-Distribution Diagnoses

AI algorithms are trained on thousands of examples. Breast cancer, prostate cancer, colon adenocarcinoma, all with abundant training data.

But what about: - Angiosarcoma (rare vascular malignancy) - Desmoplastic melanoma (melanoma mimicking benign scar tissue) - Primary bone lymphoma - Hepatosplenic T-cell lymphoma

These diagnoses may be uncommon in a given laboratory, and a development dataset may contain few or no representative cases. That must be verified rather than assumed.

What happens when AI encounters rare diagnosis?

Hypothetical failure pattern 1: confident misclassification - Algorithm trained on 10,000 benign prostate biopsies and 5,000 adenocarcinomas sees its first granulomatous prostatitis - Misclassifies the case as a common target pattern with an apparently high score - If pathologist trusts algorithm, patient gets unnecessary treatment

Hypothetical failure pattern 2: appropriate deferral - Better outcome: Algorithm recognizes it’s outside training distribution - Flags case for expert review - Whether a product detects or defers out-of-distribution inputs must be established for that product

The solution (still research-stage): - Outlier detection: Train algorithms to recognize when they’re seeing something outside training data - Confidence calibration: Algorithms should express low confidence for rare diagnoses - Human review: Follow the exact labeling and validated workflow, while retaining independent pathologist review wherever required

Current clinical approach: - Use AI only for common diagnoses where it’s been validated - Maintain high index of suspicion for rare diagnoses - When morphology, clinical context, ancillary results, and model output conflict, use an explicit escalation and diagnostic-resolution process

The page currently treats rare-diagnosis failure as a hypothetical. Park et al. tested that claim on one rare entity: 21 pathologists (7 GU specialists, 7 non-GU specialists, 7 residents) read 30 challenging H&E-only FHdRCC versus high-grade mimic cases, with and without a screening model, in a crossover design (Park et al., 2026). Median accuracy rose from 60.0% (IQR 56.7–70.0) to 73.3% (66.7–80.0; P = 0.012) and Fleiss kappa from 0.311 to 0.482; the gain was significant in GU specialists (60.0% to 76.7%, P = 0.022) but not in non-GU specialists or residents, and standalone AI accuracy on those 30 cases was 83.3% (25/30). This is an H&E screening reader study, not a deployment or outcome trial.

Part 7: Implementation Framework

Before Adopting Pathology AI

Questions to ask vendors:

  1. “What is the FDA clearance status?”
    • 510(k) cleared? PMA approved? Investigational?
    • Authorization defines a specific intended use and regulatory controls, not performance outside that scope
  2. “Was this algorithm validated on our specific scanner?”
    • Exact scanner model and software version
    • If no, plan for local validation study
  3. “What diagnoses has this been validated for?”
    • Algorithms validated for prostate cancer do not work for bladder cancer
    • Ask for published validation studies for each diagnosis
  4. “What is sensitivity and specificity for clinically significant disease?”
    • Gleason ≥7 prostate cancer (not all cancers)
    • Micrometastases ≥0.2mm (not isolated tumor cells)
    • HSIL+ cervical lesions (not ASCUS)
  5. “How does this integrate with our digital pathology workflow?”
    • PACS integration? LIS integration?
    • Does pathologist see AI annotations directly on slide viewer?
  6. “What happens when the algorithm encounters a rare diagnosis?”
    • Does it output low confidence? Flag for human review?
    • Or does it confidently misdiagnose?
  7. “Can we validate locally before clinical deployment?”
    • Define a sample and acceptance criteria appropriate to intended use and risk
    • Compare performance with a prespecified reference standard
    • Stratify by diagnosis, specimen type, staining protocol, scanner, and relevant subgroups
  8. “What is the cost structure?”
    • Per-slide fee? Annual license? Scanner-tied?
    • Hidden costs: Scanner upgrade requirements, IT integration
  9. “Who is liable if the algorithm misses a cancer?”
    • Review labeling, institutional policy, vendor terms, and jurisdiction-specific advice
    • Preserve an auditable record of the output, review, and final interpretation
  10. “Can you provide references from pathologists using this clinically?”
    • Not research collaborators. Actual clinical users
    • Ask about false positives, workflow disruptions, performance on their scanner

Red Flags (Walk Away If You See These)

  1. “Scanner-agnostic” claims without validation data: Compatibility must be demonstrated across the claimed scanners and pipelines
  2. Validated for “cancer detection” without specifying types: Each cancer requires separate validation
  3. No published peer-reviewed validation studies: Internal white papers insufficient
  4. Claims of “autonomous diagnosis”: Verify the exact labeling, user, output, and required review
  5. “Works on all staining protocols”: Require direct validation across every claimed preparation and stain condition

Part 8: Economic Evaluation

What Does Pathology AI Cost?

Prices, discounts, infrastructure arrangements, and laboratory labor change over time and are commonly confidential. The following preserved categories are the correct structure for an economic analysis, but fixed vendor prices should not be presented without a current executed proposal.

Prostate cancer detection: - Possible pricing structures include per-slide, volume-tiered, annual license, or bundled platform fees - Measure whether the authorized detection workflow changes review time, ancillary testing, discordance resolution, or diagnostic error

HER2 scoring: - AIM-HER2 is developer-described as research use only, so a diagnostic ROI claim is not established - Any reduction in confirmatory testing must remain guideline-concordant and be measured prospectively

Lymph node metastasis detection: - Not commercially available (LYNA research-only)

Cervical cytology: - Instrument, service, consumables, review time, and preparation-specific workflow belong in the model - Clinical and operational benefit must be evaluated for the exact system and target lesion threshold

Digital pathology infrastructure (prerequisite for most AI): - Whole-slide scanners, service contracts, quality control, and backup capacity - Image-management software, displays, interfaces, identity matching, and LIS integration - Network, storage, cybersecurity, archiving, disaster recovery, and data-egress costs - Training, validation, monitoring, change control, downtime, and pathologist review time

Do These Tools Save Money?

The answer is use-case and site-specific. A local analysis should compare current practice with the complete AI-enabled pathway, including downstream testing, false-positive work, missed cases, review time, implementation labor, infrastructure, monitoring, and the value assigned to clinical outcomes.

Use case Potential benefit to measure Cost or harm to measure
HER2 quantification research Reproducibility and guideline-concordant ancillary testing Validation, discordance resolution, and erroneous treatment eligibility
Prostate detection support Additional suspicious foci found, review time, ancillary stains False alerts, complete-slide review burden, integration and monitoring
Cervical cytology assistance Lesion detection, worklist prioritization, review time False-positive workload, referrals, preparation-specific failures
Lymph-node reader assistance Review time and sensitivity in the local workflow Research-to-product translation, false alerts, scanner compatibility
Digital pathology infrastructure Remote review, consultation, logistics, and platform reuse Acquisition, service, storage, interfaces, cybersecurity, validation, and downtime

The former $170,000 net-savings example combined invented prices, volumes, and a 30% testing reduction and therefore did not establish ROI. A defensible model shows low, base, and high scenarios and identifies which inputs determine the conclusion.

Part 9: Future Directions

Demonstrated, Emerging, and Speculative Directions

Demonstrated research and early clinical translation:

  1. Expanded cancer detection algorithms: Colon polyp classification, bladder cancer grading, melanoma diagnosis
  2. Molecular-feature prediction from H&E: Research models estimate selected biomarkers; clinical use still requires confirmatory strategy and prospective evidence
  3. Prognostic algorithms: Retrospective risk stratification is established in research; decision impact and patient benefit remain separate questions
  4. Workflow automation: Automated case triage, specimen tracking, quality control

Promising but uncertain:

  1. Scanner-agnostic algorithms: Stain normalization and domain adaptation to work across scanner vendors
  2. Multiplex imaging AI: Analyze multi-color immunofluorescence (tumor microenvironment, immune infiltrates)
  3. Real-time intraoperative assistance: AI analysis of frozen sections or gross images under a defined pathologist-governed workflow

Beyond the current evidence:

  1. Unbounded autonomous diagnosis across pathology: No evidence base supports reliable use across all tissues, stains, scanners, and diagnoses
  2. “One algorithm for all diagnoses”: Broad foundation-model coverage does not remove task-specific validation requirements
  3. Replacement claims: Current evidence supports bounded assistance, not elimination of pathologist-governed diagnosis

The rate-limiting factors: - Scanner standardization (biggest technical challenge) - FDA regulatory pathway clarity - Medicolegal precedent (when AI-assisted diagnosis goes wrong) - Pathologist workflow integration and trust

Key Takeaways

10 Principles for Pathology AI

  1. Pathology AI is product- and task-specific: Detection, quantification, second read, prognosis, and workflow tools have different evidence and intended uses

  2. Scanner and image-pipeline compatibility require explicit verification: Validation on one configuration does not establish performance on another

  3. Quantitative tasks require task-specific evidence: HER2 scoring, Ki-67 quantification, and mitotic count should not be grouped under one effectiveness claim

  4. Rare diagnoses require separate evidence: Common-condition performance does not establish reliable behavior on rare or unfamiliar patterns

  5. Use the labeled review model: Required pathologist review and permitted autonomy are product- and indication-specific

  6. Workflow integration determines success: Algorithm accuracy matters less than smooth PACS/LIS integration

  7. Economic claims require local evidence: Include procurement, infrastructure, confirmatory testing, staffing, and downstream care

  8. Digital pathology infrastructure is part of the intervention: Scanner, viewer, storage, network, and LIS integration affect performance and cost

  9. Demand risk-based local validation: Prespecify the sample, reference standard, subgroups, and acceptance criteria

  10. Legal questions are jurisdiction-specific: Use the canonical analysis in Liability and Malpractice

Clinical Scenario: Evaluating Prostate Cancer Detection AI

This is a hypothetical procurement exercise. The vendor, price, performance claims, department volume, and contract are illustrative and must not be read as a real product record or measured institutional result.

Scenario: Your Pathology Department Is Considering AI for Prostate Biopsies

The hypothetical pitch: A vendor demonstrates software that it says detects prostate adenocarcinoma and assigns Gleason grades. The vendor claims: - Sensitivity 98% for Gleason ≥7 cancer - “Reduces inter-pathologist variability by 30%” - FDA authorization for the proposed use - Cost: $20 per slide

For the exercise, the department processes 2,000 prostate biopsies per year. The chair asks for a recommendation.

Questions to Ask:

  1. “Was this validated on our specific scanner?”
    • We use Leica Aperio GT450. Was the algorithm trained and validated on this exact model?
    • If no, we need local validation before clinical deployment
  2. “What is performance for Gleason grade groups?”
    • Sensitivity/specificity for Grade Group 1 (3+3) vs. 2 (3+4) vs. 3+ (4+3, 4+4, 4+5)?
    • Grade Group 2 vs. 3 distinction most clinically important (active surveillance vs. treatment)
  3. “How does this handle mimics?”
    • Atypical adenomatous hyperplasia, atrophy, high-grade PIN
    • False positive rate for benign mimics?
  4. “What is the workflow integration?”
    • Integrates with our PACS (Philips IntelliSite)?
    • Do pathologists see AI annotations directly on slide viewer?
    • Adds extra clicks/steps to sign-out workflow?
  5. “Can we do local validation study?”
    • Prespecify a risk-based sample representing benign cases, relevant grade groups, mimics, scanners, stains, and difficult cases
    • Compare algorithm grades to our sign-out diagnoses
    • Stratify performance by Gleason pattern
  6. “What is the cost-benefit?”
    • 2,000 biopsies/year × $20 = $40,000/year
    • Benefits: Reduced inter-observer variability, fewer missed cancers, pathologist confidence
    • Quantifiable ROI unclear (not like HER2 where we save FISH costs)
  7. “Who is liable if algorithm misses cancer?”
    • Vendor contract liability clause?
    • Do not assume that the signer is categorically the only legally responsible party; obtain jurisdiction-specific review
  8. “How do we document AI assistance?”
    • Add statement to pathology report: “Computer-assisted detection used”?
    • Document when we disagree with algorithm?

Red Flags in This Scenario:

“Reduces variability by 30%” without baseline data: What was inter-observer variability before? Calculated how?

Authorization is summarized rather than produced: Obtain the exact FDA record and confirm specimen, scanner, software, output, intended user, and review model.

No mention of performance on Grade Group 2 vs. 3 distinction: Most clinically important classification, should be reported separately

The arithmetic is not an ROI: A per-slide fee multiplied by volume is only one cost input. The analysis also needs implementation, false alerts, ancillary testing, review time, diagnostic action, and patient-relevant benefit.

Check Your Understanding

The following three cases are hypothetical teaching exercises. They do not report real patients, real scanner failures, real product performance, or real legal outcomes.

Scenario 1: The AI-Detected Micrometastasis

Hypothetical clinical situation: You’re reviewing a sentinel lymph node from a breast cancer patient. The AI algorithm (LYNA-like system) flags a 0.3mm cluster of cells as suspicious for metastatic carcinoma. You review the flagged area. The cells are slightly enlarged with increased nuclear-to-cytoplasmic ratio, but you’re not certain they’re malignant. Could be reactive sinus histiocytosis.

Without the AI flag, you might have called this negative. With the AI flag, you’re uncertain.

Question 1: Do you sign out as “positive for micrometastasis” based on the AI flagging?

Click to reveal answer

Answer: No, not based solely on AI flagging. Get a second opinion from a breast pathology expert.

Reasoning:

Why you should not call positive based on AI alone: 1. AI can produce false alerts: The frequency and meaning depend on the study unit, threshold, prevalence, and workflow 2. Reactive histiocytes can mimic carcinoma: Enlarged cells with increased N:C ratio occur in reactive nodes 3. Clinical consequences are significant: Micrometastasis diagnosis changes staging (N0 → N1mi), may influence adjuvant therapy decisions 4. Pathologist judgment remains standard: AI is assistive, not determinative

What you should do: 1. Get expert consultation: Send digital slide to breast pathology expert for second opinion 2. Consider IHC: Cytokeratin stain (AE1/AE3 or CAM5.2) would confirm epithelial cells vs. histiocytes 3. Review clinical context: Large primary tumor, lymphovascular invasion, high grade → higher pretest probability of nodal metastasis 4. Document uncertainty: If you call it positive, note it was small focus requiring IHC confirmation or expert consultation

The AI’s value: - Flagged an area you might have missed on routine review - Prompted closer examination and consideration of IHC - But does not replace pathologist judgment for equivocal findings

Bottom line: Use AI to flag suspicious areas for closer review, but do not diagnose solely based on AI flagging when morphology is equivocal.

If IHC confirms metastatic carcinoma: AI was correct, you caught early metastasis If IHC shows histiocytes: AI was false positive, but better to overcall (with IHC confirmation) than miss metastasis

Scenario 2: The Scanner Upgrade Disaster

Hypothetical clinical situation: A department has used prostate cancer detection AI on one validated scanner configuration. Local monitoring had met prespecified acceptance criteria.

The institution changes to a different scanner model and image-processing pipeline to increase throughput.

After the scanner upgrade, you notice the AI is flagging many more false positives, with benign glands marked as suspicious. And one case where you caught a Gleason 4+3 cancer that the AI missed (it called the slide benign).

Question 2: What went wrong, and what should you do?

Click to reveal answer

Answer: Treat the change as a potential compatibility failure. Suspend reliance on the output, investigate, and follow the laboratory’s change-control and revalidation process.

What went wrong:

Scanner variability problem: 1. Different color profiles: GT450 captures RGB values differently than AT2 (different camera sensors) 2. Different compression: Image compression algorithms may differ 3. Different focus: Autofocus algorithm produces slightly different focal planes 4. Algorithm was not trained on GT450 images: Trained on AT2, does not generalize to GT450 despite being same vendor

Why this is dangerous: - You caught one false negative (Gleason 4+3 missed) - How many did you not catch? - Patients may be getting false reassurance from benign AI calls - Patient-safety and legal consequences depend on the missed finding, workflow, actions, and jurisdiction

What you should do immediately:

  1. Halt AI-assisted sign-out for clinical cases:
    • Revert to manual review without AI assistance
    • Better to lose AI benefits than risk false negatives
  2. Notify vendor:
    • Scanner upgrade broke algorithm performance
    • Request algorithm retraining/recalibration for GT450
  3. Retrospective review of cases signed out since scanner upgrade:
    • How many cases signed out with AI assistance post-upgrade?
    • Re-review all with special attention to AI-negative calls
    • Identify any missed diagnoses requiring patient notification
  4. Local acceptance or revalidation study on the new configuration:
    • Use a prespecified risk-based sample with relevant cases and mimics
    • Compare to sign-out diagnoses
    • Calculate sensitivity/specificity on new scanner
    • Only resume clinical use if performance acceptable
  5. Update validation protocols:
    • Document that algorithm is scanner-model-specific
    • Any future scanner changes require revalidation
    • Consider this in scanner procurement decisions (vendor lock-in)

Lessons learned: - Scanner specificity is real: Even upgrading within same vendor can break AI - Continuous monitoring essential: Track AI performance over time, watch for degradation - Plan for scanner changes: Budget time and resources for revalidation when upgrading scanners

Prevention for next time: - Before scanner upgrade, ask vendor: “Will this require AI revalidation?” - Plan revalidation into upgrade timeline - Consider parallel scanning (old + new scanners) during transition for validation

Scenario 3: The Rare Sarcoma Misdiagnosis

Hypothetical clinical situation: A 45-year-old woman has a breast mass biopsied. Core needle biopsy submitted. Your lab uses breast cancer detection AI that’s been excellent for usual ductal and lobular carcinomas.

The AI flags the case as “high-grade ductal carcinoma, 95% confidence.” You review the slide. The cells are indeed high-grade, pleomorphic, with mitotic activity. But the architecture does not look like typical ductal carcinoma. It’s more spindle-cell, fascicular growth pattern.

You’re not a breast pathology expert. The AI is 95% confident. Your general pathology training says “looks malignant, high-grade.”

Question 3: Do you sign out as high-grade ductal carcinoma based on AI’s 95% confidence?

Click to reveal answer

Answer: No. The morphology creates a differential that may include metaplastic carcinoma, sarcoma, melanoma, and other spindle-cell lesions. Resolve it with appropriate consultation and ancillary testing rather than accepting the model label.

Why the model output cannot resolve the case:

The rare diagnosis problem: 1. AI trained on ductal and lobular carcinomas: Seen thousands of typical breast cancers 2. Training representation is unknown: The model may have few or no examples of the relevant rare spindle-cell lesion 3. Spindle-cell pattern outside training distribution: Algorithm defaults to “high-grade carcinoma” because it’s the closest thing it knows 4. A high score may be misleading: Its mathematical meaning depends on model design and calibration. It is not automatically a 95% probability that the diagnosis is correct or a 95% similarity measure.

What the morphology requires: - A broad spindle-cell differential, not a predetermined angiosarcoma diagnosis - Correlation with architecture, clinical setting, imaging, and the complete specimen - Appropriate epithelial, vascular, melanocytic, and other ancillary markers selected by the pathologist - A final diagnosis before treatment decisions are made

What happens if you call it carcinoma: - An incorrect tumor classification can direct inappropriate staging, surgery, systemic therapy, or radiation - Patient harm from wrong diagnosis and wrong treatment

What you should do:

  1. Recognize the morphology does not fit:
    • Spindle cells ≠ ductal carcinoma
    • Fascicular pattern ≠ typical breast cancer architecture
    • Treat discordant morphology as a reason for diagnostic escalation, not as a contest between human and model confidence
  2. Get expert consultation:
    • Send to breast pathology expert at academic center
    • Describe: “High-grade spindle-cell malignancy, AI flagged as carcinoma but morphology atypical”
  3. Order IHC panel:
    • Select markers based on the differential rather than ordering a fixed panel from a vignette
    • Interpret immunohistochemistry with morphology and known limitations
    • Ancillary testing may narrow the differential but is not infallible
  4. Document AI assistance and your clinical reasoning:
    • “Computer-assisted detection system flagged as ductal carcinoma; however, spindle-cell morphology atypical for breast carcinoma. IHC and expert consultation recommended.”
  5. Report this case to AI vendor:
    • False positive (confidently misdiagnosed rare sarcoma as carcinoma)
    • Vendor should add angiosarcoma cases to training set to prevent future misdiagnoses

Lessons learned: - AI scores require calibration and task context: A displayed 95% score does not automatically mean 95% diagnostic accuracy or probability - Discordance requires resolution: When the pattern does not fit the prediction, expand the differential and use consultation or ancillary testing - Rare diagnoses are AI blind spots: Algorithms have not seen them, will misclassify as “closest common diagnosis” - Always integrate clinical context: 45-year-old woman with breast mass. Angiosarcoma is rare but recognized entity; consider it

Bottom line: Never diagnose based solely on AI prediction when morphology is atypical. AI is a tool, not truth.

Professional Society Guidance for Digital Pathology and AI

CAP Guidance Scope

The College of American Pathologists has direct guidance for whole-slide imaging and interpretive diagnostic error. These scopes should not be converted into a comprehensive AI-device validation standard.

Whole Slide Imaging (WSI) Validation Guideline (2021 online update; April 2022 publication):

Published in Archives of Pathology & Laboratory Medicine, the CAP guideline provides:

  • 3 strong recommendations (SRs)
  • 9 good practice statements (GPSs)
  • GRADE framework for evidence evaluation
  • Specific validation protocols for diagnostic use

The active CAP whole-slide imaging validation guideline addresses diagnostic accuracy and concordance of WSI systems with light microscopy. It is not a general AI-tool validation guideline.

The separate 2026 CAP Interpretive Diagnostic Error Reduction guideline addresses practices for reducing interpretive errors and changing digital environments. Its FAQ states that fixed standards for AI tools have not been established. WSI validation, laboratory-developed-test obligations, FDA device authorization, and AI-tool validation are related but distinct questions.

CAP AI Resources and Programs

Educational programs: CAP provides members educational and practice resources to:

  • Experiment with emerging AI tools in pathology
  • Explore foundation models
  • Build confidence in AI applications

Recent educational publications and resources:

  • “Introduction to Generative Artificial Intelligence: Contextualizing the Future” (February 2025)
  • “Harnessing the Power of Generative Artificial Intelligence in Pathology Education” (February 2025)
  • “Evaluating Use of Generative Artificial Intelligence in Clinical Pathology Practice” (February 2025)
  • “Bridging the Clinical-Computational Transparency Gap in Digital Pathology” (2024)

Digital Pathology Coding

Digital pathology Category III add-on codes and reporting instructions have changed over time. Laboratories should verify the current AMA CPT code set, payer policy, date of service, and whether a code describes digitization rather than AI use. A digital-pathology code is not evidence that an AI service is covered or reimbursed.

American Society for Clinical Pathology (ASCP)

ASCP participates in pathology education and collaborated on the CAP WSI guideline. The following are implementation questions retained for laboratory teams, not represented as a separate formal ASCP AI guideline:

  • Laboratory quality management for AI systems
  • Technologist training requirements for digital pathology
  • Integration of AI into laboratory workflows

Implementation framework: Laboratory implementation requires: - Validated workflows for specimen handling and digitization - Quality control procedures for scanner calibration - Clear documentation of AI involvement in diagnosis

Digital Pathology Association (DPA)

The Digital Pathology Association provides regulatory and implementation resources. These can support planning but do not replace controlling FDA records, accreditation requirements, or peer-reviewed evidence:

  • Healthcare regulatory information for digital pathology
  • Best practices for clinical deployment
  • Vendor evaluation frameworks

What is Paige Prostate Detect authorized to do?

FDA authorized Paige Prostate Detect through De Novo DEN200080 as an adjunctive aid for detecting suspicious foci on digitized prostate biopsy slides. Its authorized intended use is detection support, not Gleason scoring or grading.

Can pathology AI work across different slide scanners?

Performance can change across scanner models, software versions, stains, preparation, and image-processing pipelines. Compatibility must be demonstrated for the deployed configuration rather than assumed from a vendor name or another laboratory’s validation.

Does AI replace pathologists?

The permitted role and required human review are product- and indication-specific. Current clinical pathology AI commonly supports detection, quantification, triage, or quality control within a pathologist-governed workflow; broad autonomous-diagnosis evidence is not established.

What is pathology AI’s weakness?

Performance on common target findings does not establish performance on rare diagnoses, morphologic mimics, artifacts, or out-of-distribution tissue. Confidence scores may also be poorly calibrated outside the development distribution.

Can AI find endometrial cancer subtypes that pathologists miss?

A retrospective deep-learning study identified a p53-abnormal-like subgroup within no-specific-molecular-profile endometrial cancers and validated its adverse prognostic association in external cohorts. The study did not test model-guided treatment or establish clinical actionability.

Is AI available for gross/macroscopic pathology of surgical specimens?

Macroscopic pathology AI remains a research area. A reported lesion-localization result in one gastric-cancer dataset does not establish improved sampling or diagnostic outcomes; broader use requires standardized acquisition, external validation, and prospective workflow evidence.