Evidence Before Impression: An Agentic Systems Architecture for Multi-View Knee Radiograph Interpretation

Q: What is "evidence before impression" in radiograph AI?

"Evidence before impression" is the central design principle of 5C Network's agentic knee X-ray architecture: the system compiles structured, view-aware evidence for and against each candidate finding before any report is written, and the final report is generated only from that reduced evidence ledger. This is the opposite of one-shot image-to-report captioning, where a model writes an impression directly from pixels. By making the evidence explicit first, the report becomes auditable, the uncertainty becomes measurable, and failure modes become locatable.

Q: Why is knee X-ray interpretation hard for AI?

Knee X-rays are deceptively hard because the clinically important cases sit in the low-signal middle between clearly normal and clearly abnormal: subtle fractures, early osteoarthritis, equivocal suprapatellar effusion, projection-limited patellar findings, postoperative change, and normal variants such as the fabella that mimic pathology. A normal knee is easy. A displaced fracture is easy. The cases that matter most require multi-step, view-aware reasoning rather than a single caption.

Q: Why isn't single-pass image-to-report generation sufficient?

Unconstrained image-to-report generation is vulnerable to unsupported findings, weak projectional evidence, and poor auditability. A complete knee report requires more than a label: it requires determining which projections are available, whether each is diagnostic, which findings are best evaluated on which view, whether a suspected finding is corroborated or contradicted across views, and whether weak evidence should be reported, downgraded, or left indeterminate. When these subproblems are collapsed into one generation step, fluent prose replaces evidence-grounded interpretation.

Q: What is the evidence ledger?

The evidence ledger is a canonical structured representation of all per-finding state in a knee study. For each target finding it stores: state (present, absent, indeterminate, or not_assessable), confidence, positive evidence, negative evidence, limitations, supporting views, and grade or severity where applicable. The ledger is the system's central artifact — it is what specialists read, what the skeptic challenges, what the report writer is constrained to use, and what investigators inspect when something goes wrong. It is also a natural supervision target: an expert can correct the report, the ledger entry, or the evidence attached to it.

Q: Why four states instead of present/absent?

The four-state model — present, absent, indeterminate, not_assessable — is clinically essential. A possible effusion without a true lateral view is not equivalent to absent effusion. A suspicious lucency without a cortical break may be indeterminate rather than present. Patellar height may be not_assessable if the lateral projection is inadequate. Collapsing these distinctions into a binary present/absent label discards exactly the uncertainty a radiologist needs to act safely.

Q: What is the skeptic pass?

The skeptic pass is a finding-specific regularization layer that runs after evidence reduction. It downgrades effusion when evidence lacks direct suprapatellar pouch or capsular distention, downgrades alignment abnormality when an oblique AP projection plausibly explains apparent asymmetry, and downgrades osseous-lesion claims when the evidence is more consistent with skin, soft-tissue, or projectional artifact. It is not a generic conservatism knob — it is a set of checks grounded in radiographic view logic, designed to control false positives without suppressing true subtle findings.

Q: Is this a clinical validation study?

No. This is a technical systems manuscript prepared as scientific background for clinical validation. It describes the task formulation, system design, evidence representation, uncertainty states, trace artifacts, safety constraints, and evaluation framework — but it does not claim radiologist-level performance or report results from a multicenter reader study. 5C Network is now partnering with leading hospitals and radiology AI groups to run IRB-approved validation against radiologist-adjudicated knee cohorts.

Q: How should knee radiograph AI be evaluated?

The paper argues that knee radiograph AI should be evaluated not only by final-report accuracy, but also by view handling, evidence support, uncertainty behavior, false-positive control, clinically significant miss rate, and report signability. A rigorous design uses radiologist-adjudicated labels — independent review by two radiologists with third-reader adjudication — over a cohort that combines consecutive studies with enriched strata for low-prevalence findings, evaluated through ablations from one-shot generation up to the full ledger-constrained system.

Q: How can hospitals collaborate on IRB-approved validation?

5C Network is actively partnering with leading hospitals and radiology AI research groups to run IRB-approved validation of the agentic knee X-ray architecture. Partners can bring fully anonymised, radiologist-adjudicated knee X-ray datasets — including their own curated subtle-abnormal cohorts — for unbiased external evaluation, with structured audit artifacts retained in a controlled research environment. Collaboration formats include co-authored publications, retrospective and silent-mode prospective studies, and reader-assist studies. During evaluation the system does not replace the signed report or alter patient care. Reach out via the collaboration form on this page.

5C Applied AI Team

Our Latest Architecture · June 2026

The future of radiology AI

Evidence Before Impression

An agentic systems architecture for multi-view knee radiograph interpretation — the design behind how 5C builds radiology AI you can audit, trust, and validate.

5C Applied AI Team

5C Network · Technical systems manuscript

8 pages 7-stage reasoning graph 17-target ontology 4-state evidence model

Partner with us under IRB

Abstract

Knee radiographs are among the most common musculoskeletal imaging studies, yet automated interpretation remains challenging in precisely the cases that matter clinically: subtle fractures, early osteoarthritis, equivocal effusion, projection-limited patellar findings, and normal variants that mimic pathology.

We present an agentic systems architecture that treats reporting as structured evidence compilation rather than direct captioning — assembling multi-view studies, inferring projection roles, running task-bounded agents, reducing outputs to a canonical evidence ledger, triggering specialists when warranted, applying a skeptic pass, and synthesizing a report only from the ledger.

Technical systems manuscript · June 4, 2026 · 5C Applied AI Team

Now partnering with leading hospitals

We are accepting IRB-approved clinical trials of this architecture with leading hospitals.

Bring an anonymised, radiologist-adjudicated knee cohort. We run unbiased external evaluation, share back full traces and ledgers, and co-author the results.

Start a validation study

TL;DR: Knee X-ray AI should produce evidence before impression. Instead of captioning a study in one shot, the system assembles the multi-view study, preserves projection identity, runs task-bounded agents, reduces everything to a canonical four-state evidence ledger, triggers specialists only where warranted, applies a finding-specific skeptic pass, and writes the report only from the ledger. The contribution is not a new foundation model — it is a clinical-computational architecture that makes foundation-model reasoning inspectable, view-aware, and measurable, so it can be validated under IRB.

Published: June 4, 2026 | By the 5C Applied AI Team | Systems manuscript · 8 pages · Agentic radiology architecture

§ 01 · Motivation

Why "Evidence Before Impression"

Knee radiography is a high-volume examination across emergency medicine, orthopedics, rheumatology, sports medicine, and post-operative care. Many cases are straightforward: a normal knee may be easy to report, and a severely degenerated knee or displaced fracture may be visually obvious. The clinically important difficulty lies in the low-signal middle — early osteoarthritis, subtle avulsion fragments, tibial spine injuries, small patellar fractures, equivocal suprapatellar effusion, projection-limited alignment, and post-operative changes that can be confused with pathology.

Radiologists do not solve these cases through one-shot captioning. They first determine which projections are available and whether each is diagnostic, which findings are best evaluated on which view, whether a suspected finding is corroborated or contradicted across views, and whether weak evidence should be reported, downgraded, or left indeterminate. The final report is a reduction over evidence.

Most current AI systems do not model that workflow. CNN-based systems are effective for focused classification, but a full knee report requires more than a label. Vision-language systems can produce fluent prose, but fluency is not evidence-grounded interpretation. A credible reader must make intermediate state visible: views checked, findings considered, evidence for and against each finding, limitations, specialist decisions, and report rationale. This paper frames knee radiograph interpretation as an agentic evidence-compilation problem built for exactly that.

7

Staged steps from study assembly to ledger-constrained report

9

Task-bounded agents, each returning structured output not prose

4

States per finding: present · absent · indeterminate · not_assessable

From the paper. This is a technical systems manuscript prepared as scientific background for clinical validation — not a completed trial or a claim of radiologist equivalence.

§ 02 · Contributions

Four Contributions

The paper makes four contributions to how knee radiograph AI should be built and evaluated.

Contribution 01

Multi-view agentic architecture that preserves projection identity

A knee-specific architecture that keeps each image's projection role — AP, lateral, sunrise/Merchant, oblique — intact throughout interpretation. Projection identity is part of the diagnostic task: effusion and patellar height depend on a true lateral; compartmental narrowing on a frontal view.

Contribution 02

A structured evidence ledger for reportable clinical state

A canonical ledger that converts heterogeneous model observations into reportable clinical state — one entry per finding, with state, confidence, supporting and opposing evidence, limitations, supporting views, and grade. It is the system's central, inspectable artifact and a natural supervision target.

Contribution 03

A skeptic pass that reduces unsupported positives

A finding-specific mechanism for reducing unsupported positives arising from projectional, soft-tissue, or artifact confounds — grounded in radiographic view logic rather than generic conservatism. It downgrades effusion without capsular support, alignment overcall on oblique views, and lesion claims better explained by artifact.

Contribution 04

A validation framework beyond diagnostic labels

A framework that evaluates not only diagnostic labels but also view handling, uncertainty behavior, clinical significance of errors, and report signability — the basis for the radiologist-adjudicated, IRB-approved studies we are now running with partner hospitals.

§ 03 · Architecture

The Reasoning Graph

The system is organized as a staged reasoning graph. Each stage converts model outputs into structured intermediate state, and the final report is generated from the ledger — not directly from unconstrained image prose.

Figure 1 — The staged agentic reasoning graph. Per-view agents (03) and triggered specialists (05) call the foundation model; every other stage is deterministic control logic operating over the structured ledger.

01

Study Assembly & Ingestion

Assembles the multi-view study from DICOM, rendered radiographs, uploads, or composite screenshots. Decodes, validates, and groups images as a single study object.

02

Projection Inference

Infers projection role from metadata, filenames, and layout. A single AP/LAT composite screenshot is split into virtual per-view panels, and view identity is preserved downstream.

03

Task-Bounded Agents

Nine bounded agents — projection/coverage, rotation/obliquity, fracture/cortical integrity, alignment/dislocation, joint-space/degenerative, soft-tissue/effusion, patellar height/tracking, loose body vs fabella, postoperative/arthroplasty — each returning structured fields, not a paragraph.

04

Evidence Ledger Reduction

All agent outputs reduce into a canonical ledger Λ_k per finding: state, confidence, positive evidence, negative evidence, limitations, supporting views, grade. The central artifact for review, contradiction checks, and trace.

05

Triggered Specialist Review

Specialists run adaptively: a fracture specialist on suspicious lucency or lipohemarthrosis, an effusion specialist when lateral support is weak, a degenerative specialist on uncertain narrowing — concentrating compute where evidence is uncertain or important.

06

Skeptic Pass

Finding-specific checks grounded in view logic. Downgrades effusion without suprapatellar/capsular distention, alignment overcall explained by an oblique AP, and osseous-lesion claims better explained by skin, soft tissue, or projectional artifact.

07

Ledger-Constrained Synthesis

The report writer receives only the final ledger and writes a concise report. It cannot invent unsupported findings or re-interpret the images — otherwise the final model becomes an uncontrolled second reader.

This architecture reflects how 5C Network's Bionic AI engine is built — agentic, stateful, and inspectable rather than monolithic. Read more about Generalised Medical AI and Hybrid Intelligence.

§ 04 · Uncertainty

Four States, Not Two

For each target finding the system predicts one of four states. The distinction between genuine absence and unavailable evidence is clinically essential — and is exactly what binary present/absent labels throw away.

present

The finding is supported by direct, view-appropriate evidence.

absent

Adequately assessed and not present — distinct from simply unseen.

indeterminate

Suspicious but not corroborated — e.g. a lucency without a cortical break.

not_assessable

The required view is missing or inadequate — e.g. patellar height with no true lateral.

A possible effusion without a true lateral view is not the same as absent effusion. A suspicious lucency without a cortical break may be indeterminate rather than present. A small posterior, well-corticated ossicle may represent a fabella variant rather than a loose body. The four-state model is how the system keeps those distinctions instead of forcing a premature call.

§ 05 · Representation

The Evidence Ledger

The ledger is the system's central artifact. It enables specialist review, contradiction checks, trace inspection, and report generation — and it provides a natural supervision target: an expert can correct the report, the ledger entry, or the evidence attached to it.

Λ_k = ( s_k , q_k , E⁺_k , E⁻_k , L_k , V_k , G_k )

For each target finding y_k, the ledger stores seven fields. One row per finding, one ledger per study.

One entry, in full

ledger · study_4417 · joint_effusion SKEPTIC ACTIVE


  "finding": "joint_effusion",
  "state": "indeterminate",   // present | absent | indeterminate | not_assessable
  "confidence": 0.41,
  "evidence_for": [
     "view": "AP",  "cue": "suprapatellar soft-tissue fullness" 
  ],
  "evidence_against": [
     "view": "LAT", "cue": "no true lateral — capsular distention not assessable" 
  ],
  "limitations": ["no horizontal-beam lateral"],
  "supporting_views": ["AP"],
  "grade": null,
  "skeptic_action": "downgrade present → indeterminate (no capsular distention)"
.

A worked ledger entry. The report writer never sees the pixels — only structured rows like this one. Illustrative example, not output from a specific study.

s_k · state

present · absent · indeterminate · not_assessable

q_k · confidence

Confidence for the asserted state.

E⁺_k · positive evidence

Per-view evidence supporting the finding.

E⁻_k · negative evidence

Per-view evidence opposing the finding.

L_k · limitations

View-availability and image-quality caveats.

V_k · supporting views

Which projections support the entry.

G_k · grade

Grade or severity when applicable (e.g. OA severity).

The report writer reads only this. Traceability by construction.

§ 06 · Failure Analysis

Failure Modes the Architecture Addresses

Each recurrent failure mode in knee radiograph interpretation maps to a specific architectural mechanism. This is the table at the heart of the paper.

Failure mode	Clinical mechanism	Architectural response
Effusion overcall	Skin folds, superficial overlap, or nonspecific anterior fullness can mimic joint fluid.	Require direct lateral-view suprapatellar/capsular support; skeptic downgrade when absent.
Projectional alignment overcall	An oblique or rotated AP view can mimic varus/valgus deformity.	Track projection limitation; downgrade unsupported alignment claims.
Missed early OA	Mild spurs, tibial spine spiking, and subtle narrowing are low signal.	Degenerative specialist and ledger preservation of component evidence.
OA overgrading	Osteophytes alone can inflate severity.	Require multiple degenerative components for higher severity.
Subtle fracture miss	Small avulsion, patellar, or tibial spine findings can be normalized.	Trigger fracture specialist from direct or secondary trauma cues.
Loose body mimic	Fabella and well-corticated variants can mimic intra-articular bodies.	Separate variant/mimic target and require intra-articular support.
Hardware confusion	Post-operative changes can be normalized or misread as pathology.	Dedicated postoperative / hardware assessment.

Table 1 — Representative failure modes and corresponding system mechanisms.

§ 07 · Regularization

The Skeptic Pass

The skeptic pass regularizes weak positives. It is not a generic conservatism knob — it is a set of finding-specific checks grounded in radiographic view logic.

Effusion

Downgrade when the evidence lacks direct suprapatellar pouch or capsular distention — most often a soft-tissue or superficial-overlap confound on a view that cannot support the call.

Alignment

Downgrade alignment abnormality when an oblique AP projection plausibly explains the apparent asymmetry, rather than a true varus/valgus deformity.

Osseous lesion

Downgrade osseous-lesion claims when the evidence is more consistent with skin, soft-tissue, or projectional artifact than with a true bone lesion.

Because these checks live in the reduction layer rather than as new full model stages, they control overcall without materially increasing read time — and the same ledger discipline preserves component evidence so that genuinely subtle findings are not lost.

§ 08 · Evaluation

How It Should Be Validated

Knee radiograph AI should be evaluated not only by final-report accuracy, but by view handling, evidence support, uncertainty behavior, false-positive control, clinically significant miss rate, and report signability. This is the framework we bring to IRB-approved studies.

Reference standard

Radiologist-adjudicated labels rather than report text alone — independent review by two radiologists with third-reader adjudication for disagreements. Labels capture finding presence, view adequacy, severity, clinical significance of errors, and report signability.

Cohort design

Consecutive knee radiographs to estimate real-world operating characteristics, plus enriched strata for low-prevalence findings: subtle fracture, tibial plateau injury, patellar fracture, lipohemarthrosis, arthroplasty, loose body/fabella mimics, and poor or rotated views.

Endpoints

Primary: clinically significant miss rate and false-positive rate, fracture and effusion sensitivity/specificity, OA detection and severity calibration, and report signability. Secondary: projection-role accuracy, true-lateral identification, not-assessable appropriateness, latency, and trace completeness.

Reader-facing evaluation

A silent-mode prospective study measures operational feasibility without influencing care. A subsequent reader-assist study compares radiologist-alone and radiologist-with-system workflows — interpretation time, accepted/rejected suggestions, error correction, confidence, and automation bias.

Ablation ladder

The architecture is evaluated through five ablations that test whether each layer improves the tradeoff between sensitivity, specificity, uncertainty handling, and report signability.

1

One-shot image-to-report generation

2

Core agents without specialists

3

Core agents plus triggered specialists

4

Core agents plus specialists plus skeptic pass

5

Full ledger-constrained report synthesis

§ 09 · Governance

Clinical Safety & Governance

The system is intended for validation under appropriate institutional approval before any clinical use.

Does not replace the signed report

During evaluation the system does not replace the signed radiology report, alter patient care, or expose patient-facing outputs. All outputs are treated as research artifacts.

PHI handling follows site procedures

Protected health information handling, storage, de-identification, and access control follow site-approved data-use procedures throughout.

Clinical and research artifacts stay separate

Structured audit artifacts can be retained in a controlled research environment so radiologists and investigators inspect evidence support and failure modes — without changing the clinical report workflow.

§ 10 · Limitations

Limitations

The paper is explicit about what it does not claim.

01

A systems architecture, not a completed trial

This describes a systems architecture and validation framework, not a completed multicenter clinical trial. No claim is made that the system is ready for autonomous diagnosis.

02

Performance must be established on hospital data

Performance must be established on de-identified hospital data with radiologist adjudication and prospective silent-mode evaluation — exactly the studies we are now running with partners.

03

Decomposition has its own risks

Decomposition can propagate early errors, deterministic skeptic rules can become too conservative if poorly calibrated, and specialist routing can miss cases if triggers are too narrow. The ledger itself must be accurate.

04

Known regimes where it may fail

It may fail on rare pathology, unusual hardware, severe image-quality limitations, pediatric variants, or institution-specific acquisition patterns. No architecture removes the information limits of projection radiography.

Now accepting IRB-approved clinical trials

Trial this architecture with us under IRB

5C Network is partnering with leading hospitals and radiology AI researchers to validate the agentic knee X-ray architecture on independently curated, radiologist-adjudicated cohorts. Bring your anonymised knee X-ray dataset and we will run it as a fully unbiased external evaluation — with IRB-approved protocols, co-authored publications, silent-mode prospective studies, and reader-assist studies as the collaboration formats.

Total unbiasedness

You bring an anonymised dataset we have never seen. We evaluate on it. Results, traces, and ledgers are shared back in full.

IRB-grade rigor

Radiologist-adjudicated reference standards, IRB-approved protocols, silent-mode and reader-assist designs, and independent benchmarks — not marketing decks.

Data stays yours

Fully anonymised under DTA. No PHI. No re-use beyond the agreed study. The system never replaces your signed report during evaluation.

References

[1] Varma, M., Lu, M., Gardner, R., et al. (2019). Automated abnormality detection in lower extremity radiographs using deep learning. Nature Machine Intelligence.
[2] Antony, J., McGuinness, K., O'Connor, N. E., & Moran, K. (2016). Quantifying radiographic knee osteoarthritis severity using deep convolutional neural networks. International Conference on Pattern Recognition Workshops.
[3] Mongan, J., Moy, L., Charles, E., et al. (2024). Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 update. Radiology: Artificial Intelligence.
[4] Vasey, B., Nagendran, M., Campbell, B., et al. (2022). Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nature Medicine.
[5] Liu, X., Rivera, S. C., Moher, D., Calvert, M. J., & Denniston, A. K. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: CONSORT-AI. Nature Medicine.
[6] Rivera, S. C., Liu, X., Chan, A. W., Denniston, A. K., & Calvert, M. J. (2020). Guidelines for clinical trial protocols for interventions involving artificial intelligence: SPIRIT-AI. Nature Medicine.

PDF · 8 pages · June 2026

Read the Complete Paper

"Evidence Before Impression: An Agentic Systems Architecture for Multi-View Knee Radiograph Interpretation"

By the 5C Applied AI Team · No email required. No paywall.

Continue Reading

5C Research

All publications from the 5C Network research team.

View all papers

What is GM AI?

Generalised Medical AI — beyond narrow detection to full-workflow radiology AI.

Related Research Papers

From Slices to Reports

A survey of AI in cross-sectional medical imaging

Deep Learning for Shoulder Fracture Detection

Ensemble system for fracture identification in radiographs

Autonomous AI for Multi-Pathology Detection

Chest X-ray pathology detection across Indian hospitals

Explainable AI in Radiology

Interpretable methods for radiologist trust and adoption