In quick succession, I ran across three papers on LINKED IN on Radiology & Agentic AI.
I asked Chat GPT for a detailed summary.
Then, knowing that radiology AI is often 3-5 years ahead of pathology AI, I asked for projects from THIS work towards pathology (cf in particular digital pathology & genomics).
###
###
2026 Radiology Agentic Articles in Discussion (Plus View of
Pathology)
You’re already using something that sits on the
“agentic” spectrum: an LLM that plans, calls tools, iterates,
and persists context to finish multi-step work (your examples: complex
Medicare table work, multi-source company diligence, even finding a specific
page in a calendar PDF). The radiology literature is basically trying to (1) define
that capability precisely, (2) map where it fits in real radiology
workflows, and (3) draw safety/implementation boundaries so it doesn’t
become a flashy demo that breaks the moment it touches PACS/RIS/EHR and
liability.
Below is a detailed, paper-by-paper walk-through, plus how
it maps to your digital pathology instincts.
1) Tripathi et al. (Radiology editorial): what radiology
means by “agentic”
Tripathi/Cook/Kim are doing definitional cleanup and
systems-thinking. Their core move is to separate an AI agent (a
program—often LLM-powered—that perceives, reasons, plans, and acts via tool
calls in an iterative loop) from agentic AI (a broader systems
framework that orchestrates one or more agents/tools to solve complex,
multi-step tasks with limited human supervision).
That distinction matters because a lot of radiology “AI”
today is still narrow/static (a detector/classifier). Tripathi’s
“agentic” vision is closer to an end-to-end workflow collaborator: not
just “find nodules,” but also pull priors, query the EHR, draft
report text, check billing/documentation completeness, make
guideline-grounded follow-up suggestions, and route urgent comms—all
chained together as a goal-directed pipeline.
The practical definition (why it feels like what you’ve
seen in consulting)
Tripathi emphasizes recurring building blocks:
- Foundation
model (often LLM/MLLM) as the “brain”
- Tools/APIs
(EHR queries, guideline retrieval, image encoders, scheduling, etc.)
- Iterative
loops (observe → reflect → adjust)
- Defined
human oversight points
That’s extremely close to your “ChatGPT did elaborate
searches + calculations + document forensics” experience—just relocated into
radiology’s environment (PACS/RIS/EHR), and constrained by clinical risk.
Their “engineer’s checklist” is the most valuable part
Tripathi gets concrete about how to build/contain these
systems:
Guardrails should scale with autonomy: they
explicitly name tripwires (confidence/coverage bounds), approval
gates for consequential actions, validation checkpoints, and even execution
sandboxes.
They also insist on standards-based integration—PACS/RIS/EHR
and standards like HL7, FHIR, FHIRcast, DICOM/DICOMweb—because “agentic”
that can’t live inside the real pipes is just a lab toy.
And they highlight modern threat models that matter more for
agents than for single-purpose models: prompt injection, memory
injection, tool-surface vulnerabilities, model poisoning,
mitigated by strict I/O controls, capability whitelisting, and red-teaming.
Bottom line of Tripathi: agentic AI is a workflow
paradigm (tool-using, iterative, orchestrated), but radiology’s near-term
reality is supervised copilots, not unsupervised autonomy, because
regulation + liability + trust bar.
2) Gibson et al. (Computers, MSK scoping review): what
evidence exists, and what’s still mostly conceptual
Gibson et al. is less “here’s the grand vision” and more
“what’s actually in the literature (especially MSK), and how mature is it?”
Their helpful taxonomy: model → automated workflow →
agentic
They lay out a simple ladder:
- Single
model (e.g., CNN fracture detector)
- Automated
workflow (detector triggers alert via preset rules)
- Agentic
AI = multiple models acting as “agents” that autonomously process
information across multi-step tasks (not just “if fracture then
alert”).
That taxonomy is useful because it prevents “agentic” from
becoming a marketing synonym for “has AI somewhere.”
What they found: 11 papers total; only 2 truly
MSK-specific
Their scoping review (PubMed/Embase/Scopus/WoS) ends up with
11 included studies, with only two directly MSK-focused; the rest
are general radiology concepts or prototypes that could apply to MSK.
They explicitly characterize the evidence as limited and “largely
theoretical / exploratory.”
The four themes they extract (this is the heart of the
paper)
They organize the literature into four buckets:
- Agentic
decision support (pathway navigation, coordination, workload
reduction)
- Workflow
optimization (administrative efficiency, modality selection,
throughput)
- Image
analysis / reconstruction (multi-agent systems improving quality +
automated interpretation)
- Conceptual
/ ethical / governance (transparency, safety frameworks, clinician
oversight)
A very “your-world” detail: they emphasize “agentic” as
particularly attractive where there are structured pathways and heavy
throughput—exactly the kind of environment where orchestration and delegation
pay off.
The study table is revealing: lots of prototypes, few
validated deployments
Their Table of included studies labels maturity as things
like conceptual, prototype, ethical/legal guidance, and
only limited feasibility work.
So if radiology feels “ahead,” Gibson’s review quietly says: radiology is
ahead conceptually and in prototypes; the evidence base for real agentic
systems is still early.
Bottom line of Gibson: “agentic MSK radiology” is
mostly a map of plausible applications (decision
support/workflow/reconstruction) plus governance concerns; the field is not yet
dominated by prospective clinical trials or mature deployments.
3) Tomak (Med-Essence blog): a clean operational
contrast—RPA vs agentic AI
This piece is not academic radiology, but it’s operationally
clarifying: it draws a bright line between RPA and agentic AI in
radiology automation.
- RPA:
rule-based “software robots” mimicking clicks/keystrokes for structured
tasks (scheduling, data transfer, report distribution, billing/coding
extraction). Strength: quick deployment; limitation: brittle outside
predefined rules.
- Agentic
AI: systems that make decisions, learn/adapt, and can
handle unstructured data (images, free text), used for image
analysis, prioritization, NLP extraction, decision support. Strength:
flexibility; limitation: validation/regulatory cost and error risk without
oversight.
It’s simplistic in places (because it’s a short blog), but
it gives you a practical frame: a lot of “automation” in departments is really RPA
(valuable!), while “agentic” implies clinical-context reasoning + tool-using
orchestration, not just automating clerical steps.
Pulling the three together: what “agentic AI in
radiology” really denotes
Across Tripathi + Gibson + Tomak, “agentic” converges on:
Goal-directed orchestration of multiple steps across
systems (PACS/RIS/EHR, guidelines, priors, communication channels),
using an LLM-like planner plus tools, with iteration, memory, and
explicit human oversight gates.
That’s why your consulting examples felt “agentic”: the
system wasn’t just generating text; it was effectively acting like a junior
analyst who can decompose work, retrieve, cross-check, and
assemble.
The digital pathology translation (where you’re likely
headed)
If you port Tripathi’s radiology framing into
pathology/digital pathology, the analogous “agentic” targets aren’t just “tumor
vs non-tumor detection” but things like:
- Case
assembly agent: pull priors, relevant molecular results, previous
synoptics, outside consult PDFs, match to current specimen
- Worklist
triage agent: prioritize by clinical urgency, frozen vs routine,
re-cuts/IHC pending, critical values
- QA/compliance
agent: completeness checks (synoptic required elements, staging
fields, billing documentation), CAP checklist mapping
- Communication
agent: route criticals, generate clinician-facing summaries, ensure
acknowledgment loops
- WSI
+ text orchestration: combine slide-level signals with the chart
story, but keep approval gates and audit trails the way
Tripathi emphasizes
And your policy lens matters because regulatory/oversight
is where “agentic” will bottleneck first: the more the system acts
(orders, routes, finalizes, triggers downstream events), the more it looks like
a regulated medical device or at least a high-governance clinical decision
support layer.
A quick “so what should I take from these papers?”
summary
- Tripathi
gives the cleanest definition + architecture + guardrails: agentic
= iterative tool-using orchestration; near-term = supervised copilots; key
risks = integration, security, bias, deskilling, liability.
- Gibson
tells you the evidence base (especially MSK) is early, with lots of
conceptual/prototype work and governance discussion, not tons of mature
clinical deployments.
- Tomak
helps you keep RPA (rule automation) separate from agentic
(adaptive, unstructured, decision-support + orchestration), and suggests
hybrids will be common (RPA for admin, agentic for clinical-context
tasks).
If you want, paste (or upload) the three PDFs’
figures/tables you find most interesting (or tell me which pages), and I’ll
translate them into a pathology-forward “design pattern” memo: what to build
first, where the governance tripwires go, and what claims are
realistic vs hype—in the same pragmatic tone you use for policy work.
##
AND
VIEW TOWARD PATHOLOGY? >>>>>>>>>>>>>>>>>>>>>>>>>>
What’s portable from these radiology “agentic AI”
papers into pathology isn’t a specific algorithm—it’s the systems pattern:
an AI layer that can plan, call tools, iterate, and coordinate
across multiple clinical systems, with explicit human oversight gates.
Radiology is simply the first large specialty to articulate this as a workflow
architecture rather than a one-model-per-task story.
What the 3 radiology papers imply for pathology
Across the three papers, the biggest implications for
pathology are:
1) “Agentic” is not a better classifier; it’s workflow
orchestration.
Tripathi’s editorial treats agentic AI as a systems framework that
coordinates tools and subcomponents (LLMs, APIs, databases) to complete
multi-step clinical work. That concept ports directly to pathology because
pathology workflows are also multi-system, multi-artifact, and
latency-sensitive (case assembly, priors, orders, stains, synoptics, addenda,
billing, QA).
2) Integration is the real product.
Tripathi emphasizes that clinical viability depends on seamless integration
into existing environments and use of common interoperability standards.
Pathology’s analog isn’t PACS/RIS/EHR; it’s LIS + APLIS + digital pathology
viewer + EHR + middleware + billing/claims + QA systems. The “agent” is
often less important than the plumbing plus guardrails that make it safe
and fast.
3) Most near-term value is in “friction points,” not
headline diagnostics.
Gibson’s scoping review (especially its themes) implicitly says: the earliest
credible deployments are decision support, workflow optimization,
and quality/reconstruction, not fully autonomous interpretation. In
pathology that translates to: case triage, completeness checks, prior
retrieval, synoptic assembly, stain/workup suggestions with supervision,
specimen-to-slide chain-of-custody checks, and report distribution—before “AI
signs out cases.”
4) The RPA vs agentic distinction will matter in labs.
Tomak’s RPA-vs-agentic framing maps cleanly onto lab medicine: RPA will
keep automating structured clerical steps (orders, demographics transfer,
result routing, billing queues), while agentic AI tackles unstructured
complexity (free-text clinical history, outside reports, image regions,
longitudinal narrative, guideline-grounded recommendations). Labs will almost
certainly run hybrid stacks: RPA for throughput + agentic for cognition.
Roll the clock 3–5 years: how agentic agents “migrate”
into pathology
A plausible migration path is not “pathologists adopt
agents.” It’s “agents appear as middleware around existing systems,”
then gradually get closer to interpretation.
Phase 1 (0–18 months): agentic “case assembly and
clerical cognition”
This is the least controversial and easiest ROI:
- Case
assembler agent: pulls priors, relevant clinic notes, prior pathology,
prior molecular, imaging impressions, outside PDFs; organizes into a
single readable timeline with citations/links.
- Report
completeness agent: checks synoptic elements, staging fields, required
disclaimers, specimen/cassette counts vs blocks/slides, “did we answer the
clinical question.”
- Utilization
agent: flags redundant tests; suggests guideline-consistent
alternatives; drafts language for “why not” documentation.
- Revenue-cycle
agent: pre-screens documentation for medical necessity, coding
completeness, ABN logic, prior auth packets.
Key feature: no autonomous clinical decisions—the
agent drafts, retrieves, checks, and routes; humans approve.
Phase 2 (18–36 months): agentic “workup navigation” in
digital pathology
Now the agent starts to touch what you do next, but
still behind gates:
- Worklist
triage + routing (the radiology “triage” analog): send GI biopsies to
GI specialists, neuropath to neuropath, prioritize transplant
rejects/critical values, flag time-sensitive intraop/frozen workflows.
- Conditional
workup suggestions: “Based on pattern A + history B, consider stains
X/Y,” but requires click-to-order confirmation and
institution-specific protocols.
- Slide-level
navigation help: proposes ROIs, highlights discordant regions,
suggests “look here” with confidence estimates and “why” (so it behaves
like a junior colleague, not a black box).
- Conference
prep agent: builds tumor-board packets, links representative images,
pulls relevant molecular and guideline context, drafts a 60-second case
summary.
Phase 3 (36–60 months): bounded autonomy in narrow,
high-standardization domains
If agentic AI gets “clinical traction,” it will be in places
with high standardization, clear outcome metrics, and strong
confirmatory testing:
- Pap/urine
cytology triage, QC, rescreen logic
- IHC
scoring assistance with strict protocols and audit trails
- Frozen-section
workflow support (time management + communication + documentation; not
replacing the intraop judgment)
- High-volume
benign workflows (e.g., “negative” pre-screens) with conservative
thresholds and mandatory review
This is where Tripathi’s “human-supervised with defined
oversight points” becomes operationally decisive: the system is permitted to do
some actions, but only inside explicit boundaries.
Why it may play out differently in digital pathology vs
genomics
Both will get hit by the same tsunami, but they’ll surf
different waves.
Digital pathology: “agentic” will look like navigation +
workflow control
Digital pathology is spatial and visual. The
agent’s superpower is not just interpretation; it’s where to look, what
to do next, and how to keep the case moving through a multi-step
physical-digital chain.
Distinctive agentic opportunities in digital pathology
- Spatial
triage and attention management: ROI proposals, “second set of eyes”
on rare events, discordance detection across blocks/slides.
- Operational
intelligence: slide logistics, turn-around-time prediction, queue
balancing, stain bottleneck forecasting.
- QA
at scale: stain quality drift, scanner artifacts, tissue detection
errors, labeling mismatch detection (where available).
Distinctive constraints
- Ground
truth is messy (interobserver variability; subtle diagnostic
thresholds).
- Liability
feels closer because “the pixels are the diagnosis.”
- Standards/integration
lag: pathology interoperability is improving but still less uniform
than radiology’s PACS/DICOM culture in many settings.
So digital pathology’s agentic arc likely starts as workflow
copilots and “attention guides,” not autonomous sign-out.
Genomics: “agentic” will look like evidence curation +
reasoning over text
Genomics is symbolic, knowledge-heavy, and literature/guideline
driven. It’s almost designed for tool-using agents: search, retrieve,
compare, cite, reason, draft.
Distinctive agentic opportunities in genomics
- Variant
interpretation copilots: continuously pull ClinVar/gnomAD, primary
literature, functional data, and guideline criteria mapping (ACMG/AMP),
then draft a structured evidence table.
- Longitudinal
reinterpretation agents: monitor updates (new classifications, new
therapies, new evidence) and flag “actionable drift.”
- Report
generation with provenance: draft patient-friendly and
clinician-facing sections, with traceable citations and institution
templates.
- Coverage/prior-auth
agents: assemble payer-specific packets tying variant/test to
guidelines and policy language.
Distinctive constraints
- The
big risk isn’t “missed pixels,” it’s misstated evidence
(hallucinated citations, wrong guideline mapping).
- But
genomics is well suited to hard guardrails: retrieval-only
workflows, citation requirements, structured outputs, and evidence
checklists.
So genomics may adopt agentic systems faster because the
first wins are documentation-grade and auditable, even when
clinical interpretation remains physician-governed.
A pathology-forward “agentic” design principle you can
borrow from radiology
If you take one principle from Tripathi/Gibson and apply it
to pathology strategy, it’s this:
Start with agents that reduce cognitive load without
taking clinical authority, then gradually grant autonomy only where you can
define:
- bounded
action space (what it’s allowed to do),
- verification
gates (who must approve),
- traceability/provenance
(why it did it), and
- failure
modes (what happens when it’s uncertain or wrong).
That framework is equally relevant whether you’re building a
digital pathology agent around WSI viewers or a genomics agent around variant
knowledgebases and payer rules.
If you tell me your setting (academic vs community,
LIS/APLIS stack, and whether you already have WSI in routine sign-out), I can
map these phases into a concrete 3-year roadmap: top 5 agentic use cases,
the integration points, and the governance gates that would make
it acceptable to pathology leadership and risk/compliance.
