How Far Away are AI-Generated Pathology Reports? A Landscape and Review
##
See 22-page PDF version here.
###
Why Did We Do This Project?
FDA clearance of DeepHealth Breast Ultrasound marks an important shift from AI that merely detects abnormalities to AI that helps assemble the final diagnostic report. See K260303. See press.
New research shows how multimodal, modular, and agentic systems can integrate image findings, prior information, and structured reporting—pointing toward a radiology and pathology “diagnostic cockpit.”
How Far Away are AI-Generated Pathology
Reports? A Landscape and Review
Executive Summary
Medical
imaging artificial intelligence is moving beyond the familiar model of a
narrowly defined algorithm that detects, measures, or classifies a finding. A
newer generation of systems can assemble several steps of professional
interpretation: identify an abnormality, characterize it, integrate
measurements and prior information, and then create a preliminary diagnostic
report for specialist review.
The transition
became especially tangible in July 2026, when FDA granted 510(k) clearance to DeepHealth
Breast Ultrasound, the commercial name for the See-Mode Augmented Reporting
Tool, Breast (SMART-B; K260303).[1] The system performs
automated lesion detection and characterization and generates radiology report
findings and impressions, while leaving final assessment under radiologist
control. According to data released with the clearance, a 16-radiologist multi-reader
multi-case study found an 8% improvement in sensitivity for breast-cancer
detection and a 37% reduction in interpretation time. FDA's database confirms
K260303 as substantially equivalent. (FDA Access Data)
The scientific
literature helps explain what is emerging behind products such as SMART-B. Two
2026 breast-ultrasound studies—Azhar et al. (2026) and Huh et al.
(2026)—are particularly instructive. Rather than simply asking a
general-purpose language model to inspect an image and improvise a report, both
use modular architectures in which specialized image-analysis components
establish clinical observations and an LLM organizes those observations into
useful text. Azhar's system combines multimodal LLMs with conventional
deep-learning classifiers and report compilation; Huh's system goes further
toward an agentic “cockpit,” using an LLM as an orchestrator that invokes specialized
tools, remembers results across images, retrieves prior information, and
synthesizes the patient-level report. (PubMed)
Pathology is
following the same trajectory. HistoGPT (Tran et al., 2025) generates
dermatopathology reports directly from multiple gigapixel whole-slide images. PathChat
(Lu et al., 2024) demonstrates interactive multimodal pathology reasoning,
while Ali et al. (2026) place report generation explicitly within a
broader emerging architecture of vision-language models and multiagent systems.
Together, these developments suggest that the likely destination is not simply
an autonomous “AI pathologist.” It is a diagnostic cockpit in which
multiple AI capabilities assemble the case, examine images, retrieve context,
make measurements, draft the report, and expose evidence for pathologist
review. (Nature)
##
- The Important Transition: AI Is Beginning to Produce the Physician’s Work Product
- DeepHealth Breast Ultrasound: An FDA-Cleared Example of Analysis Feeding Report Generation
- The Academic Literature Shows What “Report Generation” Actually Means
- Azhar et al.: A Deliberately Semiautomated Reporting Pipeline
- Huh et al.: From Reporting Pipeline Toward an Agentic Diagnostic Cockpit
- Huh Also Demonstrates Why Modularity Matters
- Clinical Correctness Matters More Than Literary Similarity
- The “Cockpit” Is Already Becoming a Commercial Radiology Concept
- Aidoc First Read Suggests the Next Regulatory Step
- Radiology Research Increasingly Supports First-Draft Reporting
- Radiology Also Shows the Risks of Moving From Point Algorithms to Workflow AI
- Pathology Is Following the Same Trajectory, but the Technical Problem Is Harder
- HistoGPT Demonstrates That Whole-Slide-to-Report Generation Is Already Possible
- PathChat and Nuclei.io Show Two Other Components of the Future Pathology Cockpit
- The Most Plausible Pathology Cockpit Is a Multiagent System
- Microscopic Description and Report Prepopulation May Arrive Before Autonomous Diagnosis
- Summarization May Become Reliable Before Recommendation or Autonomous Decision-Making
- The Value Proposition Is Workflow Compression, Not Better Writing
- Safety Will Depend on Factual Correctness, Provenance, and Human Factors
- Pathology Adds a Special Layer of Risk: The Laboratory Is Part of the Input
- Regulatory Strategy: Constrain the Task, Decompose the System, Preserve Physician Control
- Digital Pathology Platforms Become the Operating System for This Future
- How Close Is AI-Generated Pathology Reporting?
- Outlook: AI May Write the Report Because It Has Already Assembled the Diagnosis
##
1. The
Important Transition: AI Is Beginning to Produce the Physician's Work Product
For most of
its clinical history, imaging AI has existed beside the physician's
interpretation. A mammography algorithm marked a suspicious lesion. A chest CT
system flagged a pulmonary embolus. A pathology algorithm quantified HER2
staining or highlighted possible tumor. The professional then incorporated the
algorithmic output into an independently constructed medical report.
Generative and
multimodal AI are beginning to change that relationship. Increasingly, the
output is not simply another score or overlay. It is part of the document
that communicates the diagnosis.
That
represents an important boundary. An algorithm reporting “87% probability of
malignancy” supplies evidence. An algorithm writing “irregular hypoechoic mass
with indistinct margins, suspicious for malignancy” has begun translating
evidence into professional interpretation. A system that then incorporates
location, dimensions, standardized classification, previous findings, and an
impression is participating in report construction.
Reporting
should not be dismissed as clerical transcription. A diagnostic report is the
final synthesis of observations, measurements, prior information,
classification rules, and professional judgment. When AI begins constructing
that synthesis, it moves from being an accessory algorithm toward becoming part
of the core diagnostic workflow.
2.
DeepHealth Breast Ultrasound: An FDA-Cleared Example of Analysis Feeding Report
Generation
The July 2026
FDA clearance of DeepHealth Breast Ultrasound is therefore unusually
important. The commercial product corresponds to the See-Mode Augmented
Reporting Tool, Breast—SMART-B—cleared under K260303. The public announcement
describes an integrated workflow encompassing automated lesion detection,
lesion characterization, and reporting. FDA's device database independently
confirms the K260303 clearance. (FDA Access Data)
The
lesion-characterization component extracts attributes including shape,
orientation, margin, echo pattern, and posterior features using terminology
aligned with ACR BI-RADS. The system also organizes measurements and findings
generated during scanning. Most consequentially for the present discussion, it generates
a radiology report containing findings and impressions, while the
interpreting radiologist retains control over the final assessment.
The workflow
can therefore be conceptualized as a continuous chain:
image
acquisition → lesion detection → lesion characterization → standardized
descriptors and measurements → findings and impression → radiologist review →
signed report.
That is
substantially more integrated than the classic CAD model.
DeepHealth
reported a multi-reader multi-case study involving 16 U.S. board-certified
radiologists. According to the FDA-submitted results released by the company,
AI assistance increased breast-cancer detection sensitivity by 8%, reduced
radiologist interpretation time by 37%, and achieved greater than 98% accuracy
in lesion localization. These pivotal SMART-B data have not yet appeared, as
far as could be identified, as a complete peer-reviewed publication. The public
materials characterize them as FDA-submitted data on file.
That makes the
eventual public 510(k) summary especially interesting. It may reveal whether
FDA treated automated reporting largely as the downstream expression of
separately validated findings, or whether it subjected report generation to
distinct clinical performance requirements. The distinction could become highly
relevant for future pathology submissions.
3. The
Academic Literature Shows What “Report Generation” Actually Means
The phrase “AI
report generation” can create a misleading mental model: an ultrasound image is
placed into an LLM and the model simply writes whatever it believes the
radiologist should say.
The current
breast-ultrasound literature describes something much more engineered.
Both Azhar
et al. (2026) and Huh et al. (2026) recognize that a real
breast-ultrasound report is assembled from multiple forms of evidence. A study
may include many images from different breast locations, B-mode images,
Doppler, shear-wave or strain elastography, measurements entered during
scanning, probe-position information, standardized BI-RADS descriptors, and
previous patient reports. (PubMed)
The problem is
therefore not principally “write some prose.” It is to associate the correct
images with the correct lesion, extract or generate the relevant observations,
preserve anatomical relationships, integrate multiple modalities, and then
assemble a patient-level document.
That
distinction has a direct parallel in pathology. A surgical pathology report
similarly may require multiple H&E slides, ancillary stains, measurements,
gross findings, previous biopsies, molecular studies, and staging rules. The
difficult part is often assembling the right diagnostic facts before writing
the sentence.
4. Azhar et
al.: A Deliberately Semiautomated Reporting Pipeline
Azhar et
al. (2026) accurately
call their system “semiautomated breast ultrasound report generation.” The
architecture deliberately retains radiologist input while automating much of
the extraction, classification, grouping, and report assembly around that
input. The study used 2,119 elastography images for development and an
independent report-generation evaluation cohort of 60 patients scanned on two
different ultrasound systems. (PubMed)
The system
contains three major modules.
The first, Image
Classification using LLM (IC-LLM), determines what kind of ultrasound image
it has received—B-mode, shear-wave elastography, strain elastography, or
Doppler. The investigators used GPT-4o mini and prompt-based visual reasoning.
Instead of requiring a manufacturer-specific classifier, the model looks for
semantic clues such as color legends, scales, labels, and units.
The second
module, Multi-Type Classification via Deep Learning (MTCDL), sends
different elastography image types to dedicated conventional deep-learning
classifiers. DenseNet-121 performed best for shear-wave imaging, with an AUROC
of 0.92, while GoogLeNet performed best for strain and Doppler imaging, with
AUROCs of 0.91 and 0.88. These are conventional image-analysis functions
embedded within the larger reporting workflow rather than replaced by the LLM.
(PubMed)
The final
module, Automatic Report Structuring and Compilation using LLM (ARSC-LLM),
is especially relevant. It extracts radiologist annotations such as lesion
dimensions and locations, combines those observations with the elastography
classifications, and groups related images according to breast side, clock-face
position, and depth. The system then builds standardized sentences from these
components.
Importantly,
the BI-RADS category remained manually entered by the radiologist. Thus the LLM
was not independently making the entire clinical diagnosis. It was functioning
largely as a semantic extraction, organization, and reporting engine
surrounding constrained image-analysis components and physician judgment.
In the
60-patient report-generation cohort, the authors reported correct generated
information for all suspicious masses evaluated, with mean processing times of
approximately 31 seconds per patient on the GE LOGIQ E10 data and 36 seconds on
the Aixplorer data. The authors explicitly recognize the system's dependence on
radiologist annotations and identify automated detection and segmentation as
potential future steps toward a more end-to-end workflow. (PubMed)
This
limitation is also instructive. The earliest clinically useful
report-generation systems do not have to automate everything. They can leave
high-risk judgments with the physician while automating much of the otherwise
fragmented work surrounding those judgments.
5. Huh et
al.: From Reporting Pipeline Toward an Agentic Diagnostic Cockpit
Huh et al.
(2026) move
conceptually farther toward the diagnostic cockpit. Their LLM is not primarily
treated as the image classifier. Instead, it functions as a manager of
specialized tools. The investigators use LangChain to coordinate several
purpose-built models and then synthesize their outputs into a report. (PubMed)
The initial
implementation includes three principal tools. A Suspicious Description Tool
uses separate neural networks to characterize lesion shape, margin, and
echogenicity. A Category Classification Tool places findings into
BI-RADS-related categories. A Probe Information Tool determines where
the image was obtained in the breast.
The system's
important feature is how those components interact. LangChain receives an
instruction, determines which tools are needed, invokes them sequentially, and
consolidates their outputs into a preliminary image-level report. It retains
those outputs in memory as subsequent images are evaluated. Once the case has
been processed, the LLM synthesizes the accumulated findings into a final
patient-level report. (PubMed)
The authors
deliberately use the term “wholistic” report generation because
conventional research often maps one image to one caption or report. Real
breast-ultrasound interpretation usually involves several images from different
locations. Huh's system attempts to automate the whole case rather than a
single image.
The framework
also adds retrieval-augmented generation. Previous patient reports can be
stored and retrieved, allowing the final output to integrate current findings
with historical information. Thus the system can assemble current image
analysis, lesion characteristics, probe location, BI-RADS categorization,
multiple preliminary reports, and relevant history into a single report.
This is
recognizably an early agentic diagnostic cockpit.
6. Huh Also
Demonstrates Why Modularity Matters
The Huh
framework illustrates another principle likely to be important for clinical AI:
the LLM does not need to be the best model at every task.
The
investigators subsequently added additional tools for suspicious-region
detection, segmentation, and optical character recognition. The overall
reporting architecture did not need to be rebuilt. The LangChain layer could
simply acquire new capabilities that could be invoked when needed.
This differs
from a monolithic end-to-end model, where introducing a new task may require
retraining much of the system.
For a future
diagnostic platform, this modularity is strategically attractive. A lesion
detector could be replaced when a better one becomes available. A measurement
module could have its own regulatory validation. An OCR tool could be upgraded
separately. A new biomarker-analysis algorithm could become another service
available to the orchestration layer.
Huh's paper
also makes clear that modularity creates its own hazard: errors can propagate.
If a location module mistakenly identifies the left breast instead of the
right, a perfectly functioning report generator can convert that error into
polished but incorrect prose. Thus modularity makes components more testable,
but it does not make the overall system automatically safe. (PubMed)
7. Clinical
Correctness Matters More Than Literary Similarity
One of the
most useful findings in Huh et al. (2026) concerns how report-generating
systems should be evaluated.
Traditional
natural-language-generation research commonly uses metrics such as BLEU,
METEOR, and ROUGE. These measure textual similarity between generated and
reference reports. Huh's system did not dominate the comparison algorithms by
these linguistic metrics.
But it
performed much better when experienced radiologists judged clinical
correctness.
Two
board-certified radiologists with 14 and 15 years of experience evaluated final
reports according to whether the system correctly represented lesion category,
probe position, and suspicious-lesion description. The proposed system achieved
a mean clinical score of 3.67 on a five-point scale and significantly
outperformed all five comparison report-generation models. Inter-rater weighted
Cohen's kappa was 0.84. (PubMed)
·
That
observation is highly relevant to pathology. Two pathology reports may be
clinically equivalent while using quite different prose.
·
Conversely,
two reports can differ by a single word—“present” versus “absent,” “positive”
versus “negative”—with enormous clinical consequences.
For medical
generative AI, factual and diagnostic fidelity is the endpoint; textual
resemblance is only a surrogate.
8. The
“Cockpit” Is Already Becoming a Commercial Radiology Concept
The cockpit
metaphor is increasingly literal. Radiology platforms are being designed as
integrated work environments rather than merely PACS viewers with separate
ancillary applications.
NewVue, for
example, explicitly markets a Radiologist Cockpit bringing worklist
functions, clinical context, prior reports, AI results, and reporting into the
same working environment. Its reporting tools are designed to draw on prior
studies, clinical history, and AI findings while supporting structured and
natural-language reporting.
NewVue — AI-assisted reporting
The important
point is not this particular vendor. It is the architecture.
An aircraft
cockpit does not contain a single “pilot algorithm.” It contains navigation,
sensors, warning systems, communications, autopilot, checklists, and displays.
The pilot's effectiveness depends partly on the integration of all of them.
Diagnostic
medicine is moving toward the same model. The specialist becomes the final
decision-maker within an increasingly intelligent workspace rather than the
person manually performing every underlying search, measurement, transcription,
and calculation.
9. Aidoc
First Read Suggests the Next Regulatory Step
Aidoc's First
Read illustrates a more explicitly generative step. FDA granted the
investigational chest-X-ray system Breakthrough Device Designation in June
2026. First Read analyzes chest radiographs and produces preliminary radiology
report drafts intended for radiologist review. It has Breakthrough
designation, not marketing clearance. (Aidoc |
Clinical AI)
Aidoc
— FDA Breakthrough Device Designation for First Read
DeepHealth and
Aidoc therefore represent different points on a spectrum. DeepHealth SMART-B
operates within a highly constrained breast-ultrasound environment, with
defined lesion attributes and standardized BI-RADS vocabulary feeding
reporting. First Read proposes broader direct interpretation and draft-report
generation for chest radiography.
If FDA
eventually clears First Read, the review could be unusually informative about
the agency's treatment of generative error, omitted findings, hallucination,
clinician oversight, and automation bias.
10.
Radiology Research Increasingly Supports First-Draft Reporting
Commercial
development is being accompanied by a rapidly strengthening peer-reviewed
evidence base.
Li et al.
(2026) developed the
Medical Imaging Report Assistant, or MIRA, using approximately 1.87 million
radiology reports from 42 hospitals across 22 Chinese provinces,
encompassing CT, MRI, and digital radiography. Twenty-four radiologists from
six centers participated in blinded evaluations. MIRA-generated impressions
were rated at least as good as reference impressions in 69% of 2,400
comparisons and reduced drafting time by approximately 0.46 minutes per report.
(RSNA Publications
Online)
Mojtahed et
al. (2026) examined an
AI tool integrated with dictation software that generated abdominal-radiology
impressions. Ten subspecialty abdominal radiologists evaluated 100 reports.
AI-generated impressions were rated equivalent or superior to
radiologist-generated impressions in 79% of evaluations for coherence, 66% for
comprehensiveness, and 77% for factual consistency. (X-ray
Interpreter)
These studies
do not establish that autonomous radiology reporting has arrived. They
establish something commercially almost as consequential: the first-draft
problem is becoming tractable.
A system does
not need to outperform a radiologist independently to create considerable
economic value. If it can reliably perform much of the routine assembly and
writing while the physician verifies and edits the result, the efficiency gain
can be substantial.
11.
Radiology Also Shows the Risks of Moving From Point Algorithms to Workflow AI
The transition
is not risk-free. Alhasan (2026) reviews radiology AI implementation
problems extending well beyond algorithm accuracy, including distribution
shift, hidden infrastructure costs, automation bias, deskilling,
interoperability, regulatory change, security, and post-deployment monitoring. (Saudi
Medical Journal)
These issues
become more serious when the AI is no longer displaying a secondary result but
actively constructing the document that the physician may sign.
A model could
be quite accurate independently yet degrade human performance if users become
anchored to confident automated suggestions. Conversely, a less-than-perfect
model could improve overall performance if its errors remain easily
recognizable while it consistently prevents omissions.
The relevant
clinical comparison therefore is not simply AI versus physician. It is
frequently physician with AI versus physician without AI.
DeepHealth's
reader-study design is interesting for precisely this reason: the company
reports changes in aided radiologist sensitivity and interpretation time rather
than relying solely on standalone algorithm accuracy.
12.
Pathology Is Following the Same Trajectory, but the Technical Problem Is Harder
Radiology
provides a useful blueprint, but pathology is not simply radiology with
different images.
A whole-slide
pathology image can contain billions of pixels. Diagnostic evidence can occur
at radically different scales—from tissue architecture at low magnification to
nuclear detail at high magnification. A case may contain several H&E
slides, multiple blocks, IHC, special stains, earlier biopsies, molecular
results, gross measurements, and clinical information.
Ali et al.
(2026) provide a
particularly useful systematic review of this emerging landscape. Their review
explicitly examines the convergence of deep learning, vision-language models,
LLMs, and multiagent systems for pathology image analysis and report
generation. They emphasize that pathology report generation differs
fundamentally from simpler image-captioning tasks because whole-slide
interpretation requires multiscale navigation, integration of image and textual
context, and mechanisms for assembling several diagnostic operations into a
coherent workflow. (Science Advances)
The inclusion
of radiology literature in Ali's review is deliberate. Radiology has moved
earlier into automated report construction and therefore provides architecture
that can be transferred into computational pathology.
[A useful
implication of Ali et al. is that the eventual pathology reporting system may
be less like “one giant model” and more like an orchestrated group of
specialist models—exactly the architecture illustrated experimentally by
Huh in breast ultrasound.]
13.
HistoGPT Demonstrates That Whole-Slide-to-Report Generation Is Already Possible
The strongest
direct evidence that pathology-report generation is no longer speculative comes
from Tran et al. (2025) and HistoGPT.
HistoGPT is a
vision-language model designed to generate dermatopathology reports from a
patient's multiple full-resolution histology images. It was trained using
15,129 whole-slide images from 6,705 patients paired with corresponding
pathology reports and evaluated across multiple external cohorts. (Nature)
Unlike
conventional image classifiers, HistoGPT does not merely return “basal cell
carcinoma” or “melanocytic nevus.” It generates report-like diagnostic text and
can also address specific clinically relevant attributes. The work demonstrates
that a model can convert information embedded in gigapixel histology into
medical-language output across a defined dermatopathology domain.
This does not
mean unrestricted autonomous pathology sign-out is ready. Performance was
strongest in more common and relatively homogeneous diagnostic categories, and
the study itself leaves considerable room for future validation and workflow
development.
But the
conceptual threshold has been crossed: whole-slide image information can be
transformed directly into a clinically recognizable pathology report.
14.
PathChat and Nuclei.io Show Two Other Components of the Future Pathology
Cockpit
Report
generation is only one component of the emerging pathology workspace.
Lu et al.
(2024) developed
PathChat as a multimodal generative AI copilot specifically for pathology.
PathChat accepts both images and text and can describe morphology, reason about
diagnoses, discuss IHC, and participate in multi-turn diagnostic dialogue. The
authors explicitly frame it as a potential human-in-the-loop diagnostic
assistant rather than an autonomous replacement for the pathologist. (Nature)
The paper is
also a useful warning. In one of the published examples, PathChat made an
erroneous statement concerning CK7/CK20 expectations in cervical carcinoma. The
authors called out the mistake themselves. That is exactly the kind of fluent
but clinically incorrect reasoning that makes physician oversight and
traceability important in generative pathology. (Nature)
A
complementary model is provided by Huang et al. (2025) and nuclei.io.
Rather than attempting broad generative diagnosis, nuclei.io uses active
learning and real-time pathologist feedback to build customized image-analysis
assistance. Stanford describes it as a human-in-the-loop framework that learns
which cells or structures an individual pathologist wants highlighted. (Stanford
Medicine)
Together,
these systems illustrate different cockpit functions: one provides
conversational reasoning and synthesis; another provides adaptive image
navigation and detection. A future integrated platform could contain both.
15. The
Most Plausible Pathology Cockpit Is a Multiagent System
A mature
pathology AI system is unlikely to consist of one enormous model looking at all
the slides and improvising an answer.
The
architecture described by Ali et al. (2026) is more plausible:
specialized AI agents perform different tasks and exchange information. One may
analyze image features; another retrieves clinical history; another produces a
preliminary description; another checks terminology, staging rules, or medical
ontologies; and another assembles the final report. (Science Advances)
The concept is
already visible in Huh's working breast-ultrasound implementation. The LLM's
most important role is not necessarily “being smarter” than every specialized
algorithm. Its role is knowing which tool to call, what the tool's answer
means, what information remains missing, and how to assemble the pieces.
A pathology
implementation could therefore include slide-quality control, tumor detection,
tumor classification, quantitative measurement, IHC interpretation, prior-case
retrieval, molecular-result retrieval, staging logic, report construction, and
final consistency checking.
ü
The
visible output would be prose.
ü
The
real product would be the orchestration underneath the prose.
16.
Microscopic Description and Report Prepopulation May Arrive Before Autonomous
Diagnosis
The road to
AI-generated pathology reports does not require a single regulatory leap from
digital microscopy to autonomous AI sign-out.
Reporting
functions can be separated according to clinical risk.
·
At
the lower end, software can prepopulate specimen identifiers, measurements,
previous diagnoses, and structured fields. It can transform
pathologist-confirmed observations into a microscopic description. It can
insert synoptic elements once the underlying findings are established.
·
A
higher level would generate diagnostic wording from image findings. Higher
still would be automated grading, staging, or differential diagnosis. The most
consequential systems would decide which ancillary tests should be ordered or
make management-relevant recommendations.
These
functions need not arrive simultaneously or receive identical regulatory
treatment.
This suggests
that microscopic-description drafting may be one of the lowest-hanging
fruits in pathology. If the pathologist has already selected the relevant
region and confirmed the essential diagnosis, AI can assume much of the
report-writing burden without having to carry independent responsibility for
the entire diagnostic process.
HistoGPT's
work supports this possibility, while the Azhar architecture provides a
regulatory analogy: retain physician control over the highest-level clinical
judgment while automating much of the data extraction and report assembly
surrounding it.
17.
Summarization May Become Reliable Before Recommendation or Autonomous
Decision-Making
Evidence
outside radiology and pathology reinforces this progression.
Yim et al.
(2026) evaluated LLMs
for tumor-board case summarization and treatment recommendation across multiple
institutions and model families. Clinicians generally rated AI-generated case
summaries substantially better than AI-generated recommendations. Summary
ratings ranged approximately 3.57–4.59 on a five-point scale, while
recommendation-generation scores were roughly 2.0–3.6. (Sciety)
That asymmetry
is revealing.
AI may become
excellent at assembling the relevant facts, organizing them, summarizing them,
and creating a useful first draft before it becomes equally trustworthy
at independently deciding what should be done.
The probable
pathology sequence may therefore be:
case
assembly and data retrieval → report prepopulation → microscopic-description
drafting → structured diagnostic synthesis → differential diagnosis and
ancillary-test suggestions → increasingly autonomous diagnosis.
The first
several stages are considerably closer than the last.
18. The
Value Proposition Is Workflow Compression, Not Better Writing
The commercial
rationale for generative reporting should not be reduced to saving dictation
time.
A pathology
report is the final product of dozens of small workflow transactions. The
pathologist examines several slides, finds the previous biopsy, checks the
gross measurement, reviews stains, looks up molecular results, calculates
staging elements, completes synoptic fields, formulates the diagnosis, dictates
the report, and proofreads the final text.
A mature
cockpit can compress many of those activities into one coordinated process.
Before the
pathologist finishes looking at the case, AI could already have identified
tumor-containing slides, retrieved the prior specimen, collected relevant
biomarkers, calculated measurements, populated structured fields, checked
staging criteria, and highlighted unresolved discrepancies.
The report
then becomes almost the natural consequence of the assembled case.
This is why
Huh's work is conceptually important. The LangChain system's most interesting
innovation is not its prose. It is its case memory and orchestration,
which make the prose possible.
The long-term
value of AI-generated pathology reports may therefore reside less in “writing”
and more in eliminating the friction between all the steps that precede
writing.
19. Safety
Will Depend on Factual Correctness, Provenance, and Human Factors
Generative
reporting creates a distinctive safety problem: a fluent error can look
completely professional.
A poor
segmentation may be visibly poor. A suspiciously high risk score may invite
review. An LLM can take an incorrect intermediate result and transform it into
a polished sentence that sounds authoritative.
For this
reason, conventional NLP similarity measures cannot serve as the primary safety
endpoint. Huh et al. provide an unusually useful empirical demonstration: their
system did not lead all conventional language metrics, but radiologists judged
its clinical content substantially better than competing report-generation
approaches. (PubMed)
Pathology
validation should therefore ask direct factual questions. Did the system omit
an invasive focus? Did it invent lymphovascular invasion? Did it confuse the
right and left specimens? Did it change a measurement? Did it convert
“negative” into “positive”? Did it assign the wrong grade or stage?
Provenance may become equally important. A
generated statement such as “lymphovascular invasion present” should ideally
link directly to the supporting image region. A margin measurement should link
to the image and measurement that produced it. A biomarker statement should
link to the relevant assay result.
The cockpit
therefore should expose evidence rather than merely output conclusions.
20.
Pathology Adds a Special Layer of Risk: The Laboratory Is Part of the Input
A pathology AI
system does not simply receive an image from a standardized imaging device. It
receives the downstream output of an entire laboratory process.
Fixation
can vary. Processing can vary. Section thickness can vary. H&E staining can
vary. IHC can vary. Scanner optics and compression can vary. Artifacts can
differ among laboratories.
Thus, the
performance chain is:
specimen
→ fixation → processing → sectioning → staining → scanning → image analysis →
diagnostic reasoning → report generation.
Errors or
distribution shifts introduced anywhere upstream can propagate into the final
narrative.
This is one
reason digital-pathology infrastructure becomes more important as generative
capabilities increase. Scanner metadata, QC information, stain information,
specimen identity, and laboratory workflow state may need to become inputs to
the AI itself rather than invisible background conditions.
Ali et al.
emphasize generalizability, multidevice variability, interoperability, and the
scarcity of sufficiently representative real-world validation as central
barriers to translation. (PubMed
Central (PMC))
21.
Regulatory Strategy: Constrain the Task, Decompose the System, Preserve
Physician Control
The DeepHealth
SMART-B clearance suggests an attractive early regulatory pattern.
The clinical
domain is narrow. The modality is defined. The reporting vocabulary is highly
standardized. Intermediate image-analysis outputs are constrained. The
reporting function operates downstream from those outputs. A qualified
physician reviews the final assessment.
That pattern
could translate naturally into pathology.
A
plausible first-generation pathology submission might involve one organ, one
defined specimen type, one image preparation, a limited diagnostic space,
specified measurements and classifications, structured preliminary reporting,
and mandatory pathologist review.
This is
considerably more regulatable than a product labeled simply “AI software for
generating pathology diagnoses and reports.”
A modular
architecture may help further. If lesion detection, measurement, grading, and
report generation can be separately evaluated, both regulators and laboratories
can understand where errors arise.
In that sense,
Azhar and Huh may represent two useful regulatory archetypes. Azhar
keeps the highest-level clinical assessment clearly with the radiologist while
automating surrounding work. Huh demonstrates how increasingly
autonomous specialist tools can be orchestrated while remaining separately
identifiable and replaceable.
The likely
development path is therefore constrain first, validate individual
functions, and generalize gradually.
22. Digital
Pathology Platforms Become the Operating System for This Future
If the cockpit
model is correct, digital-pathology infrastructure becomes strategically more
valuable.
The platform
is no longer merely where a pathologist views a whole-slide image. It becomes
the operating environment through which multiple diagnostic services interact.
That
environment eventually needs access to whole-slide images, scanner metadata,
the LIS, gross pathology, previous specimens, special stains, IHC, molecular
results, radiology, relevant EHR data, image-analysis algorithms, staging
rules, structured reporting, and the signed report.
Storage and
compute therefore become part of the clinical architecture. So do
interoperability, auditability, latency, cybersecurity, and orchestration.
A superior
standalone algorithm that requires manual exporting and re-entry may be less
useful than a slightly weaker model that is deeply integrated into the case
workflow.
Radiology is
already demonstrating this transition. The growing “cockpit” concept treats
clinical context, algorithms, prior imaging, reporting, and workflow as a
single environment.
Digital
pathology is likely to follow.
23. How
Close Is AI-Generated Pathology Reporting?
Several
components are already technically close.
·
Report
prepopulation and structured data entry require little speculative technology.
·
Microscopic-description
drafting appears
feasible now, particularly when the pathologist has already supplied or
confirmed the diagnosis.
·
Bounded
image-to-report generation for common diseases has already been demonstrated
experimentally by HistoGPT.
·
Interactive
multimodal diagnostic assistance
has been demonstrated by PathChat.
·
Human-customized
image-navigation assistance
has been demonstrated by nuclei.io.
And multiagent
report-generation architectures now have both a substantial conceptual
literature, summarized by Ali et al., and practical analogues in other imaging
domains such as Huh's breast-ultrasound system.
What remains
substantially harder is broad autonomous surgical-pathology reporting. Such a
system must reliably search enormous slides, integrate multiple specimens and
stains, recognize rare disease, determine when material is inadequate,
reconcile clinical and molecular context, and know when a case falls outside
its validated domain.
These are not
principally writing problems.
The language
technology is increasingly adequate.
The difficult
problem is constructing a reliable diagnostic system that earns the right to
write the report.
24.
Outlook: AI May Write the Report Because It Has Already Assembled the Diagnosis
The DeepHealth
clearance makes 2026 look increasingly like an inflection point. FDA has now
cleared a system that moves from image analysis through standardized
characterization into report generation in a physician-supervised workflow. (FDA Access Data)
Aidoc is
pursuing a broader chest-X-ray model in which AI directly creates a preliminary
report draft. (Aidoc |
Clinical AI)
Azhar shows
how multimodal LLMs and conventional deep learning can automate image
classification, information extraction, and structured report assembly while
preserving radiologist control. Huh shows the next conceptual step: an LLM
acting as an orchestrator of specialized models, retaining case-level memory,
retrieving prior information, and creating the final report. (PubMed)
Pathology now
has its own corresponding developments. HistoGPT demonstrates
whole-slide-to-report generation. PathChat demonstrates multimodal diagnostic
conversation. Nuclei.io demonstrates adaptive human-in-the-loop image
assistance. Most importantly for the larger architecture, Ali et al. (2026)
explicitly describe the convergence of pathology image analysis,
vision-language models, LLMs, and multiagent systems into integrated diagnostic
and reporting workflows. (Science Advances)
The likely
endpoint is therefore not well summarized as “AI writes; doctor signs.”
The deeper
transformation is that AI may increasingly perform much of the work that
happens before the writing. It will retrieve the history, find the
important tissue, call specialized algorithms, make measurements, reconcile
results, identify missing information, check required elements, and organize
the diagnostic evidence.
Once that work
has been performed, report generation becomes almost the inevitable final
output.
That is why
the diagnostic cockpit is a more useful model than either “AI assistant”
or “autonomous pathologist.” The specialist remains in command, but an
increasing fraction of the navigation, retrieval, calculation, documentation,
and cross-checking occurs inside a coordinated computational environment.
·
Radiology
is already entering that environment.
·
For
digital pathology, it may define the next generation of the field.
|
Sidebar: How Are AI-Generated Radiology and Pathology
Reports Different?
AI-generated
radiology and pathology reports are converging toward the same broad idea—AI
analyzes images, assembles relevant findings, and drafts a report for
specialist review—but the underlying technical problems are quite
different. Radiology
is the easier starting point. A chest X-ray, CT series, MRI, or
ultrasound study is already natively digital, produced by standardized
imaging equipment, and usually associated with a well-defined anatomical exam
and reporting vocabulary. In breast ultrasound, for example, systems can
detect a lesion, characterize shape, margin, orientation, and echogenicity,
assign or assist with BI-RADS descriptors, and then assemble these elements
into structured findings and impressions. The Huh and Azhar studies show that
radiology report generation can be organized around discrete image-level
tools, prior studies, probe position, and standardized reporting frameworks.
Pathology
is fundamentally more complex because the “image” is really a case composed
of many specimens and information sources. A single pathology
case may contain numerous gigapixel whole-slide images, multiple blocks,
H&E slides, immunohistochemistry, special stains, gross measurements,
prior biopsies, molecular results, and clinical history. The diagnostic
system must first find the relevant tissue across enormous images and
multiple magnifications before it can begin composing prose. HistoGPT
demonstrates that direct whole-slide-to-report generation is possible in
bounded domains, but broad surgical pathology remains substantially harder. The
likely difference, therefore, is one of degree of orchestration.
Radiology report generation can often begin with image interpretation and
structured report assembly. Pathology will require a larger “diagnostic
cockpit” that coordinates slide navigation, image analysis, ancillary
studies, prior-case retrieval, staging rules, and evidence provenance before
the report is drafted.
In
both fields, the near-term model is not autonomous sign-out. It is AI-generated
first draft plus specialist verification. The crucial distinction is that
pathology AI must solve a much larger case-assembly problem before it earns
the right to write the report.
|
References
Ali U,
Shafi I, Ahmad J, et al. 2026.
A Systematic Literature Review on Integrated Deep Learning and Multiagent
Vision-Language Frameworks for Pathology Image Analysis and Report Generation.
Computational and Structural Biotechnology Journal. 35:0023.
Ali et al. — full article / DOI
(Science Advances)
Alhasan MS.
2026. Artificial
Intelligence in Radiology: Hidden Fragilities and the Path to Resilience.
Saudi Medical Journal. 47:195–218.
Alhasan — full
article (Saudi
Medical Journal)
Azhar K,
Lee B-D, Byon SS, Lee S, Cho KR, Song SE. 2026. Semiautomated breast ultrasound
report generation using multimodal large language models and deep learning.
Frontiers in Medicine. 13:1679203. PMID 41647521.
Azhar et al. — PubMed
Azhar et al. — open-access
article (PubMed)
Huang Z, et
al. 2025. A
pathologist-AI collaboration framework for enhancing diagnostic accuracy and
efficiency.
Huang et al. — PubMed (PubMed)
Huh J, Ahn
HS, Park HJ, Ye JC. 2026.
Wholistic report generation for Breast ultrasound using LangChain. Computerized
Medical Imaging and Graphics. 127:102697. PMID 41477975.
Huh et al. — PubMed
Huh et al. —
article DOI (PubMed)
Li M, Wang
Y, Miao Z, et al. 2026.
Fine-Tuned Large Language Model for Automated Radiology Impression
Generation: A Multicenter Evaluation. Radiology: Artificial Intelligence.
8:e250714. PMID 41983921.
Li et al. — PubMed
Li et al. — RSNA article (RSNA Publications
Online)
Lu MY, Chen
B, Williamson DFK, et al. 2024.
A multimodal generative AI copilot for human pathology. Nature.
634:466–473.
Lu et al. —
PathChat, Nature (Nature)
Mojtahed A,
Ghosh S, Goiffon RJ, et al. 2026.
Automated generation of impressions in abdominal radiology reports using an
artificial intelligence-based tool: performance compared to manual impressions.
Abdominal Radiology.
Mojtahed et al. — DOI
(X-ray
Interpreter)
Tran M,
Schmidle P, Guo RR, et al. 2025.
Generating dermatopathology reports from gigapixel whole slide images with
HistoGPT. Nature Communications. 16:4886. PMID 40419470.
Tran et al. —
HistoGPT, Nature Communications
Tran et al. — PubMed (Nature)
Yim W-W,
Damm H, Pakull TMG, et al. 2026.
Challenges in AI Based Tumor Board Case Summarization and Recommendations.
Preprint posted June 16, 2026.
Yim et al. — DOI/preprint
(Sciety)
Regulatory
and Industry Sources
DeepHealth
/ See-Mode SMART-B, K260303.
FDA-cleared breast-ultrasound software combining lesion detection,
characterization, and automated reporting.
FDA
— See-Mode K260303 device listing (FDA Access Data)
Aidoc First
Read. FDA Breakthrough
Device Designation for AI-assisted chest-X-ray interpretation and preliminary
report drafting.
Aidoc
— First Read announcement (Aidoc |
Clinical AI)
NewVue
Radiologist Cockpit.
Commercial example of the integrated diagnostic-workspace model.
NewVue — Radiologist Cockpit
NewVue — Reporting
Stanford
Medicine — nuclei.io.
Human-in-the-loop pathology AI and workflow assistance.
Stanford
Medicine — AI tool gives pathologists speed, accuracy and a new way to
collaborate (Stanford
Medicine)
##
